bump 1.4.0

+ added wikitools
+ Improved webpage reading tool

README.md

@@ -3,7 +3,7 @@
   <br />
 
   <!-- Logo -->
-  <img src="https://git.zakscode.com/repo-avatars/a90851ca730480ec37a5c0c2c4f1b4609eee5eadf806eaf16c83ac4cb7493aa9" alt="Logo" width="200" height="200">
+  <img alt="Logo" width="200" height="200" src="https://git.zakscode.com/repo-avatars/a82d423674763e7a0c1c945bdbb07e249b2bb786d3c9beae76d5b196a10f5c0f">
 
   <!-- Title -->
 ### @ztimson/ai-utils
@@ -53,13 +53,15 @@ A TypeScript library that provides a unified interface for working with multiple
 - **Provider Abstraction**: Switch between AI providers without changing your code
 
 ### Built With
-[![Anthropic](https://img.shields.io/badge/Anthropic-191919?style=for-the-badge&logo=anthropic&logoColor=white)](https://anthropic.com/)
+[![Anthropic](https://img.shields.io/badge/Anthropic-de7356?style=for-the-badge&logo=anthropic&logoColor=white)](https://anthropic.com/)
-[![OpenAI](https://img.shields.io/badge/OpenAI-412991?style=for-the-badge&logo=openai&logoColor=white)](https://openai.com/)
+[![llama](https://img.shields.io/badge/llama.cpp-fff?style=for-the-badge&logo=ollama&logoColor=black)](https://github.com/ggml-org/llama.cpp)
-[![Ollama](https://img.shields.io/badge/Ollama-000000?style=for-the-badge&logo=ollama&logoColor=white)](https://ollama.com/)
+[![OpenAI](https://img.shields.io/badge/OpenAI-000?style=for-the-badge&logo=openai-gym&logoColor=white)](https://openai.com/)
-[![TensorFlow](https://img.shields.io/badge/TensorFlow-FF6F00?style=for-the-badge&logo=tensorflow&logoColor=white)](https://tensorflow.org/)
+[![Pyannote](https://img.shields.io/badge/Pyannote-458864?style=for-the-badge&logo=python&logoColor=white)](https://github.com/pyannote)
-[![Tesseract](https://img.shields.io/badge/Tesseract-3C8FC7?style=for-the-badge&logo=tesseract&logoColor=white)](https://tesseract-ocr.github.io/)
+[![TensorFlow](https://img.shields.io/badge/TensorFlow-fff?style=for-the-badge&logo=tensorflow&logoColor=ff6f00)](https://tensorflow.org/)
+[![Tesseract](https://img.shields.io/badge/Tesseract-B874B2?style=for-the-badge&logo=hack-the-box&logoColor=white)](https://tesseract-ocr.github.io/)
+[![Transformers.js](https://img.shields.io/badge/Transformers.js-000?style=for-the-badge&logo=hugging-face&logoColor=yellow)](https://huggingface.co/docs/transformers.js/en/index)
 [![TypeScript](https://img.shields.io/badge/TypeScript-3178C6?style=for-the-badge&logo=typescript&logoColor=white)](https://typescriptlang.org/)
-[![Whisper](https://img.shields.io/badge/Whisper-412991?style=for-the-badge&logo=openai&logoColor=white)](https://github.com/ggerganov/whisper.cpp)
+[![Whisper](https://img.shields.io/badge/Whisper.cpp-000?style=for-the-badge&logo=openai-gym&logoColor=white)](https://github.com/ggerganov/whisper.cpp)
 
 ## Setup
 
@@ -75,6 +77,7 @@ A TypeScript library that provides a unified interface for working with multiple
 
 #### Instructions
 1. Install the package: `npm i @ztimson/ai-utils`
+2. For speaker diarization: `pip install pyannote.audio`
 
 </details>
 
@@ -87,17 +90,138 @@ A TypeScript library that provides a unified interface for working with multiple
 
 #### Prerequisites
 - [Node.js](https://nodejs.org/en/download)
+- _[Whisper.cpp](https://github.com/ggml-org/whisper.cpp/releases/tag) (ASR)_
+- _[Pyannote](https://github.com/pyannote) (ASR Diarization):_ `pip install pyannote.audio`
 
 #### Instructions
 1. Install the dependencies: `npm i`
-2. Build library: `npm build`
+2. For speaker diarization: `pip install pyannote.audio`
-3. Run unit tests: `npm test`
+3. Build library: `npm build`
+4. Run unit tests: `npm test`
 
 </details>
 
 ## Documentation
 
-[Available Here](https://ai-utils.docs.zakscode.com/)
+### Setup
+```javascript
+const ai = new Ai({
+    path: '/ai-models',
+    
+    // Setup audio
+    whisper: '/path/to/binary', // Required for ASR
+    hfToken: '...', // Required for diarization
+    asr: 'ggml-base.en.bin', // Override default ASR model
+    
+    // Setup LLM
+    embedder: 'bge-small-en-v1.5', // Override default embedder model
+    llm: {
+        system: 'You are a helpful assistant.',
+        compress: {max: 90_000, min: 50_000}, // Compress chat history to min tokens when max is reached
+        temperature: 0.8,
+        max_tokens: 100_000,
+        memoryModel: 'gpt-4o', // Cheap model for managing memories in background, defaults to current model
+        models: {
+            'claude-3-5-sonnet': {proto: 'anthropic', token: process.env.ANTHROPIC_TOKEN},
+            'gpt-4o':            {proto: 'openai',    token: process.env.OPENAI_TOKEN},
+            'llama3':            {proto: 'ollama',    host: 'http://localhost:11434'},
+        },
+        mcp: [
+            {name: 'files', url: 'https://mcp.example.com', token: process.env.MCP_TOKEN}
+        ],
+        skills: [
+            {name: 'Tone of voice', description: 'Brand writing guidelines', content: '# Tone of Voice\n\nAlways be concise and friendly...'}
+        ],
+        tools: [{
+            name: 'Marco?',
+            description: 'Where is marco polo?',
+            args: {
+                shout: {type: 'boolean', default: 'Shout into the void?', description: false, required: false}
+            },
+            fn: (args: any, stream: LLMRequest['stream'], ai: Ai) => {
+                const {shout} = args;
+                return shout ? 'Polo!' : 'Polo';
+            }
+        }],
+    },
+	
+    // Setup Vision
+    ocr: 'eng' // Override default OCR model
+});
+
+```
+
+### Audio
+
+```javascript
+// Crate audio transcript
+const text = await ai.audio.asr('./path/to/audio.mp3');
+console.log(text);
+
+// Break transcript into speakers
+const text = await ai.audio.asr('./path/to/audio.mp3', {diarization: true});
+console.log(text);
+
+// Break transcript into named speakers
+const text = await ai.audio.asr('./path/to/audio.mp3', {diarization: 'llm'});
+console.log(text);
+```
+
+### Language
+
+```javascript
+const history = [], memory = [];
+
+// Wait for entire response
+const text = await ai.language.ask('My favorite color is blue, whats yours?', {history, memory});
+console.log(text);
+
+// Stream response
+const chunks = '';
+await ai.language.ask('Write me a poem', {
+	history, memory,
+	stream: chunk => chunks += chunk,
+});
+console.log(chunks);
+
+// Manually compile history into memories at end of conversation
+// Happens automatically when coverstaions are compressed
+await ai.language.updateMemory(history, memory);
+
+// Summarize text
+const summary = await ai.language.summarize(longText, 200);
+
+// Code response (no conversation or extra BS)
+const code = await ai.language.code('Write a fibonacci function');
+
+// Structured JSON response
+const data = await ai.language.json('Extract the name and age', `{
+    "name": "string",
+    "age": "number"
+}`, {system: 'Extract from user input'});
+```
+
+#### Premade LLM Tools:
+- `cli`: Run a shell command, returns its output
+- `get_datetime`: Returns local date/time
+- `get_datetime_utc`: Returns current UTC date/time
+- `exec`: Execute code in cli, node, or python
+- `fetch`: Make HTTP requests (GET/POST/PUT/DELETE)
+- `exec_javascript`: Execute CommonJS JavaScript
+- `exec_python`: Execute Python via python -c
+- `read_webpage`: Scrape & clean content from a URL, handles HTML, JSON, CSV, media, PDFs etc.
+- `web_search`: Anonymous DuckDuckGo search, returns a list of URLs
+- `wikipedia_lookup`: Fetch a Wikipedia article (intro or full)
+- `wikipedia_search`: Search Wikipedia and return matching articles
+- `get_weather`: Fetch current weather + forecast for a location (just built!)
+
+### Vision
+
+```javascript
+// Extract text from image
+const text = await ai.vision.ocr('./path/to/image.png');
+console.log(text);
+```
 
 ## License
 

package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@ztimson/ai-utils",
-	"version": "0.1.5",
+	"version": "1.4.0",
 	"description": "AI Utility library",
 	"author": "Zak Timson",
 	"license": "MIT",
@@ -16,7 +16,7 @@
 		".": {
 			"types": "./dist/index.d.ts",
 			"import": "./dist/index.mjs",
-			"require": "./dist/index.cjs"
+			"require": "./dist/index.js"
 		}
 	},
 	"scripts": {
@@ -25,20 +25,21 @@
 		"watch": "npx vite build --watch"
 	},
 	"dependencies": {
-		"@anthropic-ai/sdk": "^0.67.0",
+		"@anthropic-ai/sdk": "^0.102.0",
 		"@tensorflow/tfjs": "^4.22.0",
-		"@ztimson/node-utils": "^1.0.4",
+		"@huggingface/transformers": "^4.2.0",
-		"@ztimson/utils": "^0.27.9",
+		"@ztimson/node-utils": "^1.0.7",
-		"ollama": "^0.6.0",
+		"@ztimson/utils": "^0.29.4",
-		"openai": "^6.6.0",
+		"cheerio": "^1.2.0",
-		"tesseract.js": "^6.0.1"
+		"openai": "^6.42.0",
+		"tesseract.js": "^7.0.0"
 	},
 	"devDependencies": {
-		"@types/node": "^24.8.1",
+		"@types/node": "^24.13.1",
 		"typedoc": "^0.26.7",
-		"typescript": "^5.3.3",
+		"typescript": "^5.6.3",
-		"vite": "^5.0.12",
+		"vite": "^8.0.16",
-		"vite-plugin-dts": "^4.5.3"
+		"vite-plugin-dts": "^5.0.2"
 	},
 	"files": [
 		"dist"

src/ai.ts

@@ -1,115 +1,44 @@
-import {$} from '@ztimson/node-utils';
+import * as os from 'node:os';
-import {createWorker} from 'tesseract.js';
+import LLM, {AnthropicConfig, OpenAiConfig, LLMRequest} from './llm';
-import {LLM, LLMOptions} from './llm';
+import { Audio } from './audio.ts';
-import fs from 'node:fs/promises';
+import {Vision} from './vision.ts';
-import Path from 'node:path';
-import * as tf from '@tensorflow/tfjs';
 
-export type AiOptions = LLMOptions & {
+export type AbortablePromise<T> = Promise<T> & {
-	whisper?: {
+	abort: () => any
-		/** Whisper binary location */
+};
-		binary: string;
+
-		/** Model */
+export type AiOptions = {
-		model: WhisperModel;
+	/** Token to pull diarization models from hugging face */
-		/** Working directory for models and temporary files */
+	hfToken?: string;
-		path: string;
+	/** Path to models */
+	path?: string;
+	/** Whisper ASR model: ggml-tiny.en.bin, ggml-base.en.bin */
+	asr?: string;
+	/** Embedding model: all-MiniLM-L6-v2, bge-small-en-v1.5, bge-large-en-v1.5 */
+	embedder?: string;
+	/** Large language models, first is default */
+	llm?: Omit<LLMRequest, 'model'> & {
+		models: {[model: string]: AnthropicConfig | OpenAiConfig};
 	}
+	/** OCR model: eng, eng_best, eng_fast */
+	ocr?: string;
+	/** Whisper binary */
+	whisper?: string;
 }
 
-export type WhisperModel = 'tiny' | 'base' | 'small' | 'medium' | 'large';
-
 export class Ai {
-	private downloads: {[key: string]: Promise<void>} = {};
+	/** Audio processing AI */
-	private whisperModel!: string;
+	audio!: Audio;
-
+	/** Language processing AI */
-	/** Large Language Models */
+	language!: LLM;
-	llm!: LLM;
+	/** Vision processing AI */
+	vision!: Vision;
 
 	constructor(public readonly options: AiOptions) {
-		this.llm = new LLM(this, options);
+		if(!options.path) options.path = os.tmpdir();
-		if(this.options.whisper?.binary) this.downloadAsrModel(this.options.whisper.model);
+		process.env.TRANSFORMERS_CACHE = options.path;
-	}
+		this.audio = new Audio(this);
-
+		this.language = new LLM(this);
-	/**
+		this.vision = new Vision(this);
-	 * Convert audio to text using Auditory Speech Recognition
-	 * @param {string} path Path to audio
-	 * @param model Whisper model
-	 * @returns {Promise<any>} Extracted text
-	 */
-	async asr(path: string, model?: WhisperModel): Promise<string | null> {
-		if(!this.options.whisper?.binary) throw new Error('Whisper not configured');
-		if(!model) model = this.options.whisper.model;
-		await this.downloadAsrModel(<string>model);
-		const name = Math.random().toString(36).substring(2, 10) + '-' + path.split('/').pop();
-		const output = Path.join(this.options.whisper.path || '/tmp', name);
-		await $`rm -f /tmp/${name}.txt && ${this.options.whisper.binary} -nt -np -m ${this.whisperModel} -f ${path} -otxt -of ${output}`;
-		return fs.readFile(output, 'utf-8').then(text => text?.trim() || null)
-			.finally(() => fs.rm(output, {force: true}).catch(() => {}));
-	}
-
-	/**
-	 * Downloads the specified Whisper model if it is not already present locally.
-	 *
-	 * @param {string} model Whisper model that will be downloaded
-	 * @return {Promise<void>} A promise that resolves once the model is downloaded and saved locally.
-	 */
-	async downloadAsrModel(model: string): Promise<void> {
-		if(!this.options.whisper?.binary) throw new Error('Whisper not configured');
-		this.whisperModel = Path.join(<string>this.options.whisper?.path, this.options.whisper?.model + '.bin');
-		if(await fs.stat(this.whisperModel).then(() => true).catch(() => false)) return;
-		if(!!this.downloads[model]) return this.downloads[model];
-		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${this.options.whisper?.model}.bin`)
-			.then(resp => resp.arrayBuffer()).then(arr => Buffer.from(arr)).then(async buffer => {
-				await fs.writeFile(this.whisperModel, buffer);
-				delete this.downloads[model];
-			});
-		return this.downloads[model];
-	}
-
-	/**
-	 * Convert image to text using Optical Character Recognition
-	 * @param {string} path Path to image
-	 * @returns {{abort: Function, response: Promise<string | null>}} Abort function & Promise of extracted text
-	 */
-	ocr(path: string): {abort: () => void, response: Promise<string | null>} {
-		let worker: any;
-		return {
-			abort: () => { worker?.terminate(); },
-			response: new Promise(async res => {
-				worker = await createWorker('eng');
-				const {data} = await worker.recognize(path);
-				await worker.terminate();
-				res(data.text.trim() || null);
-			})
-		}
-	}
-
-	/**
-	 * Compare the difference between two strings using tensor math
-	 * @param target Text that will checked
-	 * @param {string} searchTerms Multiple search terms to check against target
-	 * @returns {{avg: number, max: number, similarities: number[]}} Similarity values 0-1: 0 = unique, 1 = identical
-	 */
-	semanticSimilarity(target: string, ...searchTerms: string[]) {
-		if(searchTerms.length < 2) throw new Error('Requires at least 2 strings to compare');
-
-		const vector = (text: string, dimensions: number = 10): number[] => {
-			return text.toLowerCase().split('').map((char, index) =>
-				(char.charCodeAt(0) * (index + 1)) % dimensions / dimensions).slice(0, dimensions);
-		}
-
-		const cosineSimilarity = (v1: number[], v2: number[]): number => {
-			if (v1.length !== v2.length) throw new Error('Vectors must be same length');
-			const tensor1 = tf.tensor1d(v1), tensor2 = tf.tensor1d(v2)
-			const dotProduct = tf.dot(tensor1, tensor2)
-			const magnitude1 = tf.norm(tensor1)
-			const magnitude2 = tf.norm(tensor2)
-			if(magnitude1.dataSync()[0] === 0 || magnitude2.dataSync()[0] === 0) return 0
-			return dotProduct.dataSync()[0] / (magnitude1.dataSync()[0] * magnitude2.dataSync()[0])
-		}
-
-		const v = vector(target);
-		const similarities = searchTerms.map(t => vector(t)).map(refVector => cosineSimilarity(v, refVector))
-		return {avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length, max: Math.max(...similarities), similarities}
 	}
 }

src/antrhopic.ts

@@ -1,36 +1,49 @@
 import {Anthropic as anthropic} from '@anthropic-ai/sdk';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, makeArray} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
+import {TokenPool} from './token-pool.ts';
+import {convertSchema} from './tools.ts';
 
 export class Anthropic extends LLMProvider {
-	client!: anthropic;
+	private clients = new Map<string, anthropic>();
+	tokenPool!: TokenPool;
 
-	constructor(public readonly ai: Ai, public readonly apiToken: string, public model: string) {
+	constructor(public readonly ai: Ai, public readonly apiToken: string | string[], public model: string) {
 		super();
-		this.client = new anthropic({apiKey: apiToken});
+		this.tokenPool = new TokenPool(...makeArray(apiToken).filter(Boolean));
+	}
+
+	private getClient(token: string): anthropic {
+		let client = this.clients.get(token);
+		if(!client) {
+			client = new anthropic({apiKey: token});
+			this.clients.set(token, client);
+		}
+		return client;
 	}
 
 	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
+		const timestamp = Date.now();
-			const orgI = i;
+		const messages: LLMMessage[] = [];
-			if(typeof history[orgI].content != 'string') {
+		for(let h of history) {
-				if(history[orgI].role == 'assistant') {
+			if(typeof h.content == 'string') {
-					history[orgI].content.filter((c: any) => c.type =='tool_use').forEach((c: any) => {
+				messages.push(<any>{timestamp, ...h});
-						i++;
+			} else {
-						history.splice(i, 0, {role: 'tool', id: c.id, name: c.name, args: c.input});
+				const textContent = h.content?.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
-					});
+				if(textContent) messages.push({role: h.role, content: textContent, timestamp: timestamp, duration: h.duration, tps: h.tps});
-				} else if(history[orgI].role == 'user') {
+				h.content.forEach((c: any) => {
-					history[orgI].content.filter((c: any) => c.type =='tool_result').forEach((c: any) => {
+					if(c.type == 'tool_use') {
-						const h = history.find((h: any) => h.id == c.tool_use_id);
+						messages.push({role: 'tool', id: c.id, name: c.name, args: c.input, timestamp: h.timestamp, content: undefined, duration: h.duration, tps: h.tps});
-						h[c.is_error ? 'error' : 'content'] = c.content;
+					} else if(c.type == 'tool_result') {
-					});
+						const m: any = messages.findLast(m => (<any>m).id == c.tool_use_id);
-				}
+						if(m) m[c.is_error ? 'error' : 'content'] = c.content;
-				history[orgI].content = history[orgI].content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
+					}
+				});
 			}
 		}
-		return history.filter(h => !!h.content);
+		return messages;
 	}
 
 	private fromStandard(history: LLMMessage[]): any[] {
@@ -47,17 +60,20 @@ export class Anthropic extends LLMProvider {
 		return history;
 	}
 
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
+	ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
 		const controller = new AbortController();
-		const response = new Promise<any>(async (res, rej) => {
+		return Object.assign(new Promise<any>(async (res) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
+			let history = this.fromStandard([
-			if(options.compress) history = await this.ai.llm.compress(<any>history, options.compress.max, options.compress.min, options);
+				...(options.history || []).filter(h => h.role !== 'system'),
+				{role: 'user', content: message, timestamp: Date.now()}
+			]);
+			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
-				max_tokens: options.max_tokens || this.ai.options.max_tokens || 4096,
+				max_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || 4096,
-				system: options.system || this.ai.options.system || '',
+				system: options.system || this.ai.options.llm?.system || '',
-				temperature: options.temperature || this.ai.options.temperature || 0.7,
+				temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
-				tools: (options.tools || this.ai.options.tools || []).map(t => ({
+				tools: tools.map(t => ({
 					name: t.name,
 					description: t.description,
 					input_schema: {
@@ -71,12 +87,26 @@ export class Anthropic extends LLMProvider {
 				stream: !!options.stream,
 			};
 
-			// Run tool changes
+			// Add structured output support
-			let resp: any;
+			if(options.schema) {
-			do {
+				requestParams.output_config = {
-				resp = await this.client.messages.create(requestParams);
+					format: {
+						type: 'json_schema',
+						schema: convertSchema(options.schema)
+					}
+				};
+			}
 
-				// Streaming mode
+			let resp: any, terminal = false, duration = 0, tps = 0;
+			do {
+				requestParams.messages = history.map(({timestamp, ...m}) => m);
+				const callStart = Date.now();
+				resp = await this.tokenPool.run(token => this.getClient(token).messages.create(requestParams)).catch(err => {
+					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
+					throw err;
+				});
+
+				let usage: any;
 				if(options.stream) {
 					resp.content = [];
 					for await (const chunk of resp) {
@@ -97,37 +127,58 @@ export class Anthropic extends LLMProvider {
 							}
 						} else if(chunk.type === 'content_block_stop') {
 							const last = resp.content.at(-1);
-							if(last.input != null) last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
+							if(last?.input != null) last.input = last.input ? JSONAttemptParse(last.input, {}) : {};
+						} else if(chunk.type === 'message_delta') {
+							if(chunk.usage) usage = chunk.usage;
 						} else if(chunk.type === 'message_stop') {
 							break;
 						}
 					}
+				} else {
+					usage = resp.usage;
 				}
+				duration = Date.now() - callStart;
+				tps = usage?.output_tokens && duration > 0 ? usage.output_tokens / (duration / 1000) : 0;
 
-				// Run tools
 				const toolCalls = resp.content.filter((c: any) => c.type === 'tool_use');
 				if(toolCalls.length && !controller.signal.aborted) {
-					history.push({role: 'assistant', content: resp.content});
+					history.push({role: 'assistant', content: resp.content, timestamp: Date.now(), duration, tps});
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
-						const tool = options.tools?.find(findByProp('name', toolCall.name));
+						const tool = tools.find(findByProp('name', toolCall.name));
+						if(options.stream) options.stream({tool: toolCall.name});
 						if(!tool) return {tool_use_id: toolCall.id, is_error: true, content: 'Tool not found'};
 						try {
-							const result = await tool.fn(toolCall.input, this.ai);
+							const toolStream = options.stream && ((chunk: any) => {
-							return {type: 'tool_result', tool_use_id: toolCall.id, content: JSONSanitize(result)};
+								if(chunk.done) { terminal = true; return; }
+								options.stream!(chunk);
+							});
+							const result = await tool.fn(toolCall.input, toolStream, this.ai, toolCall.id);
+							return {type: 'tool_result', tool_use_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result};
 						} catch (err: any) {
 							return {type: 'tool_result', tool_use_id: toolCall.id, is_error: true, content: err?.message || err?.toString() || 'Unknown'};
 						}
 					}));
-					history.push({role: 'user', content: results});
+					history.push({role: 'user', content: results, timestamp: Date.now()});
 					requestParams.messages = history;
 				}
-			} while (!controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
+			} while (!terminal && !controller.signal.aborted && resp.content.some((c: any) => c.type === 'tool_use'));
+
+			if(!terminal) {
+				const textContent = resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n');
+				history.push({role: 'assistant', content: textContent.trim(), timestamp: Date.now(), duration, tps});
+			}
+
+			history = this.toStandard(history);
+			if(options.history) options.history.splice(0, options.history.length, ...history);
 			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {
+
-				role: 'assistant',
+			const turnStart = history.map(h => h.role).lastIndexOf('user');
-				content: resp.content.filter((c: any) => c.type == 'text').map((c: any) => c.text).join('\n\n')
+			const finalContent = history.slice(turnStart + 1).reduce((str, h) => {
-			}]));
+				if(h.role === 'assistant') return str + (h.content || '');
-		});
+				return str;
-		return Object.assign(response, {abort: () => controller.abort()});
+			}, '').trim();
+
+			res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
+		}), {abort: () => controller.abort()});
 	}
 }

src/audio.ts

@@ -0,0 +1,276 @@
+import {execSync, spawn} from 'node:child_process';
+import {mkdtempSync} from 'node:fs';
+import fs from 'node:fs/promises';
+import {tmpdir} from 'node:os';
+import Path, {join} from 'node:path';
+import {AbortablePromise, Ai} from './ai.ts';
+
+export class Audio {
+	private downloads: {[key: string]: Promise<string>} = {};
+	private pyannote!: string;
+	private whisperModel!: string;
+
+	constructor(private ai: Ai) {
+		if(ai.options.whisper) {
+			this.whisperModel = ai.options.asr || 'ggml-base.en.bin';
+			this.downloadAsrModel();
+		}
+
+		this.pyannote = `
+import sys
+import json
+import os
+from pyannote.audio import Pipeline
+
+os.environ['TORCH_HOME'] = r"${ai.options.path}"
+pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", token="${ai.options.hfToken}")
+output = pipeline(sys.argv[1])
+
+segments = []
+for turn, speaker in output.speaker_diarization:
+    segments.append({"start": turn.start, "end": turn.end, "speaker": speaker})
+
+print(json.dumps(segments))
+`;
+	}
+
+	private async addPunctuation(timestampData: any, llm?: boolean, cadence = 150): Promise<string> {
+		const countSyllables = (word: string): number => {
+			word = word.toLowerCase().replace(/[^a-z]/g, '');
+			if(word.length <= 3) return 1;
+			const matches = word.match(/[aeiouy]+/g);
+			let count = matches ? matches.length : 1;
+			if(word.endsWith('e')) count--;
+			return Math.max(1, count);
+		};
+
+		let result = '';
+		timestampData.transcription.filter((word, i) => {
+			let skip = false;
+			const prevWord = timestampData.transcription[i - 1];
+			const nextWord = timestampData.transcription[i + 1];
+			if(!word.text && nextWord) {
+				nextWord.offsets.from = word.offsets.from;
+				nextWord.timestamps.from = word.offsets.from;
+			} else if(word.text && word.text[0] != ' ' && prevWord) {
+				prevWord.offsets.to = word.offsets.to;
+				prevWord.timestamps.to = word.timestamps.to;
+				prevWord.text += word.text;
+				skip = true;
+			}
+			return !!word.text && !skip;
+		}).forEach((word: any) => {
+			const capital = /^[A-Z]/.test(word.text.trim());
+			const length = word.offsets.to - word.offsets.from;
+			const syllables = countSyllables(word.text.trim());
+			const expected = syllables * cadence;
+			if(capital && length > expected * 2 && word.text[0] == ' ') result += '.';
+			result += word.text;
+		});
+		if(!llm) return result.trim();
+		return this.ai.language.ask(result, {
+			system: 'Remove any misplaced punctuation from the following ASR transcript using the replace tool. Avoid modifying words unless there is an obvious typo',
+			temperature: 0.1,
+			tools: [{
+				name: 'replace',
+				description: 'Use find and replace to fix errors',
+				args: {
+					find: {type: 'string', description: 'Text to find', required: true},
+					replace: {type: 'string', description: 'Text to replace', required: true}
+				},
+				fn: (args) => result = result.replace(args.find, args.replace)
+			}]
+		}).then(() => result);
+	}
+
+	private async diarizeTranscript(timestampData: any, speakers: any[], llm: boolean): Promise<string> {
+		const speakerMap = new Map();
+		let speakerCount = 0;
+		speakers.forEach((seg: any) => {
+			if(!speakerMap.has(seg.speaker)) speakerMap.set(seg.speaker, ++speakerCount);
+		});
+
+		const punctuatedText = await this.addPunctuation(timestampData, llm);
+		const sentences = punctuatedText.match(/[^.!?]+[.!?]+/g) || [punctuatedText];
+		const words = timestampData.transcription.filter((w: any) => w.text.trim());
+
+		// Assign speaker to each sentence
+		const sentencesWithSpeakers = sentences.map(sentence => {
+			sentence = sentence.trim();
+			if(!sentence) return null;
+
+			const sentenceWords = sentence.toLowerCase().replace(/[^\w\s]/g, '').split(/\s+/);
+			const speakerWordCount = new Map<number, number>();
+
+			sentenceWords.forEach(sw => {
+				const word = words.find((w: any) => sw === w.text.trim().toLowerCase().replace(/[^\w]/g, ''));
+				if(!word) return;
+
+				const wordTime = word.offsets.from / 1000;
+				const speaker = speakers.find((seg: any) => wordTime >= seg.start && wordTime <= seg.end);
+				if(speaker) {
+					const spkNum = speakerMap.get(speaker.speaker);
+					speakerWordCount.set(spkNum, (speakerWordCount.get(spkNum) || 0) + 1);
+				}
+			});
+
+			let bestSpeaker = 1;
+			let maxWords = 0;
+			speakerWordCount.forEach((count, speaker) => {
+				if(count > maxWords) {
+					maxWords = count;
+					bestSpeaker = speaker;
+				}
+			});
+
+			return {speaker: bestSpeaker, text: sentence};
+		}).filter(s => s !== null);
+
+		// Merge adjacent sentences from same speaker
+		const merged: Array<{speaker: number, text: string}> = [];
+		sentencesWithSpeakers.forEach(item => {
+			const last = merged[merged.length - 1];
+			if(last && last.speaker === item.speaker) {
+				last.text += ' ' + item.text;
+			} else {
+				merged.push({...item});
+			}
+		});
+
+		let transcript = merged.map(item => `[Speaker ${item.speaker}]: ${item.text}`).join('\n').trim();
+		if(!llm) return transcript;
+		let chunks = this.ai.language.chunk(transcript, 500, 0);
+		if(chunks.length > 4) chunks = [...chunks.slice(0, 3), <string>chunks.at(-1)];
+		await this.ai.language.ask(chunks.join('\n'), {
+			system: 'Read the following transcript and attempt to identify every speaker. For every positively identified speaker, call the \`identify\` tool with the speaker\'s ID number & the identified name exactly once.',
+			temperature: 0.1,
+			tools: [
+				{name: 'identify', description: 'Identify a speaker', args: {
+					speaker: {type: 'number', description: 'Speaker number', required: true},
+					name: {type: 'string', description: 'Inferred name', required: true},
+				}, fn: ({speaker, name}) => {
+					transcript = transcript.replaceAll(`[Speaker ${speaker}]`, `[${name}]`);
+				}}
+			]
+		});
+		return transcript;
+	}
+
+	private runAsr(file: string, opts: {model?: string, diarization?: boolean} = {}): AbortablePromise<any> {
+		let proc: any;
+		const p = new Promise<any>((resolve, reject) => {
+			this.downloadAsrModel(opts.model).then(m => {
+				if(opts.diarization) {
+					let output = join(Path.dirname(file), 'transcript');
+					proc = spawn(<string>this.ai.options.whisper,
+						['-m', m, '-f', file, '-np', '-ml', '1', '-oj', '-of', output],
+						{stdio: ['ignore', 'ignore', 'pipe']}
+					);
+					proc.on('error', (err: Error) => reject(err));
+					proc.on('close', async (code: number) => {
+						if(code === 0) {
+							output = await fs.readFile(output + '.json', 'utf-8');
+							fs.rm(output + '.json').catch(() => { });
+							try { resolve(JSON.parse(output)); }
+							catch(e) { reject(new Error('Failed to parse whisper JSON')); }
+						} else {
+							reject(new Error(`Exit code ${code}`));
+						}
+					});
+				} else {
+					let output = '';
+					proc = spawn(<string>this.ai.options.whisper, ['-m', m, '-f', file, '-np', '-nt']);
+					proc.on('error', (err: Error) => reject(err));
+					proc.stdout.on('data', (data: Buffer) => output += data.toString());
+					proc.on('close', async (code: number) => {
+						if(code === 0) {
+							resolve(output.trim() || null);
+						} else {
+							reject(new Error(`Exit code ${code}`));
+						}
+					});
+				}
+			});
+		});
+		return <any>Object.assign(p, {abort: () => proc?.kill('SIGTERM')});
+	}
+
+	private runDiarization(file: string): AbortablePromise<any> {
+		let aborted = false, abort = () => { aborted = true; };
+		const checkPython = (cmd: string) => {
+			return new Promise<boolean>((resolve) => {
+				const proc = spawn(cmd, ['-W', 'ignore', '-c', 'import pyannote.audio']);
+				proc.on('close', (code: number) => resolve(code === 0));
+				proc.on('error', () => resolve(false));
+			});
+		};
+		const p = Promise.all<any>([
+			checkPython('python'),
+			checkPython('python3'),
+		]).then(<any>(async ([p, p3]: [boolean, boolean]) => {
+			if(aborted) return;
+			if(!p && !p3) throw new Error('Pyannote is not installed: pip install pyannote.audio');
+			const binary = p3 ? 'python3' : 'python';
+			return new Promise((resolve, reject) => {
+				if(aborted) return;
+				let output = '';
+				const proc = spawn(binary, ['-W', 'ignore', '-c', this.pyannote, file]);
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.stderr.on('data', (data: Buffer) => console.error(data.toString()));
+				proc.on('close', (code: number) => {
+					if(code === 0) {
+						try { resolve(JSON.parse(output)); }
+						catch (err) { reject(new Error('Failed to parse diarization output')); }
+					} else {
+						reject(new Error(`Python process exited with code ${code}`));
+					}
+				});
+				proc.on('error', reject);
+				abort = () => proc.kill('SIGTERM');
+			});
+		}));
+		return <any>Object.assign(p, {abort});
+	}
+
+	asr(path: string, options: { model?: string; diarization?: boolean | 'llm' } = {}): AbortablePromise<string | null> {
+		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
+
+		const tmp = join(mkdtempSync(join(tmpdir(), 'audio-')), 'converted.wav');
+		execSync(`ffmpeg -i "${path}" -ar 16000 -ac 1 -f wav "${tmp}"`, { stdio: 'ignore' });
+		const clean = () => fs.rm(Path.dirname(tmp), {recursive: true, force: true}).catch(() => {});
+
+		if(!options.diarization) return this.runAsr(tmp, {model: options.model});
+		const timestamps = this.runAsr(tmp, {model: options.model, diarization: true});
+		const diarization = this.runDiarization(tmp);
+		let aborted = false, abort = () => {
+			aborted = true;
+			timestamps.abort();
+			diarization.abort();
+			clean();
+		};
+
+		const response = Promise.allSettled([timestamps, diarization]).then(async ([ts, d]) => {
+			if(ts.status == 'rejected') throw new Error('Whisper.cpp timestamps:\n' + ts.reason);
+			if(d.status == 'rejected') throw new Error('Pyannote:\n' + d.reason);
+			if(aborted || !options.diarization) return ts.value;
+			return this.diarizeTranscript(ts.value, d.value, options.diarization == 'llm');
+		}).finally(() => clean());
+		return <any>Object.assign(response, {abort});
+	}
+
+	async downloadAsrModel(model: string = this.whisperModel): Promise<string> {
+		if(!this.ai.options.whisper) throw new Error('Whisper not configured');
+		if(!model.endsWith('.bin')) model += '.bin';
+		const p = Path.join(<string>this.ai.options.path, model);
+		if(await fs.stat(p).then(() => true).catch(() => false)) return p;
+		if(!!this.downloads[model]) return this.downloads[model];
+		this.downloads[model] = fetch(`https://huggingface.co/ggerganov/whisper.cpp/resolve/main/${model}`)
+			.then(resp => resp.arrayBuffer())
+			.then(arr => Buffer.from(arr)).then(async buffer => {
+				await fs.writeFile(p, buffer);
+				delete this.downloads[model];
+				return p;
+			});
+		return this.downloads[model];
+	}
+}

src/embedder.ts

@@ -0,0 +1,13 @@
+import { pipeline } from '@huggingface/transformers';
+
+const [modelDir, model] = process.argv.slice(2);
+
+let text = '';
+process.stdin.on('data', chunk => text += chunk);
+process.stdin.on('end', async () => {
+	const embedder = await pipeline('feature-extraction', 'Xenova/' + model, {cache_dir: modelDir});
+	const output = await embedder(text, { pooling: 'mean', normalize: true });
+	const embedding = Array.from(output.data);
+	process.stdout.write(JSON.stringify({embedding}));
+	process.exit();
+});

src/index.ts

@@ -1,4 +1,10 @@
 export * from './ai';
 export * from './antrhopic';
+export * from './audio';
 export * from './llm';
+export * from './memory';
+export * from './open-ai';
+export * from './provider';
+export * from './token-pool'
 export * from './tools';
+export * from './vision';

src/kd-tree.ts

@@ -0,0 +1,334 @@
+export type DistanceMetric = "euclidean" | "cosine";
+
+export interface KDPoint<T = unknown> {
+	vector: number[];
+	payload: T;
+}
+
+export interface KNNResult<T = unknown> {
+	point: KDPoint<T>;
+	distance: number;
+}
+
+interface KDNode<T> {
+	point: KDPoint<T>;
+	axis: number;
+	left: KDNode<T> | null;
+	right: KDNode<T> | null;
+}
+
+// ─── Distance helpers ─────────────────────────────────────────────────────────
+
+function euclidean(a: number[], b: number[]): number {
+	let sum = 0;
+	for (let i = 0; i < a.length; i++) {
+		const d = a[i] - b[i];
+		sum += d * d;
+	}
+	return Math.sqrt(sum);
+}
+
+function cosine(a: number[], b: number[]): number {
+	let dot = 0, normA = 0, normB = 0;
+	for (let i = 0; i < a.length; i++) {
+		dot   += a[i] * b[i];
+		normA += a[i] * a[i];
+		normB += b[i] * b[i];
+	}
+	const denom = Math.sqrt(normA) * Math.sqrt(normB);
+	return denom === 0 ? 1 : 1 - dot / denom; // distance = 1 - similarity
+}
+
+/**
+ * Keeps the k closest candidates in memory, evicts the furthest when full
+ */
+class BoundedMaxHeap<T> {
+	private heap: KNNResult<T>[] = [];
+
+	constructor(private readonly k: number) {}
+
+	get size(): number { return this.heap.length; }
+
+	get worstDistance(): number {
+		return this.heap.length < this.k ? Infinity : this.heap[0].distance;
+	}
+
+	push(item: KNNResult<T>): void {
+		if (this.heap.length < this.k) {
+			this.heap.push(item);
+			this.bubbleUp(this.heap.length - 1);
+		} else if (item.distance < this.heap[0].distance) {
+			this.heap[0] = item;
+			this.sinkDown(0);
+		}
+	}
+
+	toSortedArray(): KNNResult<T>[] {
+		return [...this.heap].sort((a, b) => a.distance - b.distance);
+	}
+
+	private bubbleUp(i: number): void {
+		while (i > 0) {
+			const parent = (i - 1) >> 1;
+			if (this.heap[parent].distance >= this.heap[i].distance) break;
+			[this.heap[parent], this.heap[i]] = [this.heap[i], this.heap[parent]];
+			i = parent;
+		}
+	}
+
+	private sinkDown(i: number): void {
+		const n = this.heap.length;
+		while (true) {
+			let largest = i;
+			const l = 2 * i + 1, r = 2 * i + 2;
+			if (l < n && this.heap[l].distance > this.heap[largest].distance) largest = l;
+			if (r < n && this.heap[r].distance > this.heap[largest].distance) largest = r;
+			if (largest === i) break;
+			[this.heap[largest], this.heap[i]] = [this.heap[i], this.heap[largest]];
+			i = largest;
+		}
+	}
+}
+
+/**
+ * K-D Tree for efficient nearest-neighbor search over high-dimensional vectors / embeddings.
+ *
+ * Supports:
+ *  - Insertion of labeled points
+ *  - k-nearest-neighbor (KNN) search
+ *  - Radius search (all points within a given distance)
+ *  - Euclidean and cosine distance metrics
+ *  - Bulk construction (balanced tree) for best query performance
+ */
+export class KDTree<T = unknown> {
+	private root: KDNode<T> | null = null;
+	private _size = 0;
+	private readonly dims: number;
+	private readonly distanceFn: (a: number[], b: number[]) => number;
+
+	/**
+	 * @param dims           Dimensionality of all vectors (must be consistent).
+	 * @param metric         Distance metric to use. Default: "euclidean".
+	 * @param points         Optional initial set of points. Builds a balanced tree
+	 *                       in O(n log² n) — prefer this over inserting one-by-one
+	 *                       when you have a large corpus.
+	 */
+	constructor(
+		dims: number,
+		metric: DistanceMetric = "euclidean",
+		points?: KDPoint<T>[]
+	) {
+		this.dims = dims;
+		this.distanceFn = metric === "cosine" ? cosine : euclidean;
+
+		if (points && points.length > 0) {
+			this.validateAll(points);
+			this.root = this.buildBalanced([...points], 0);
+			this._size = points.length;
+		}
+	}
+
+	/** Total number of points stored in the tree. */
+	get size(): number { return this._size; }
+
+	// ── Insertion ──────────────────────────────────────────────────────────────
+
+	/**
+	 * Insert a single point. O(log n) average, O(n) worst case on skewed data.
+	 * For bulk loading prefer passing points to the constructor.
+	 */
+	insert(point: KDPoint<T>): void {
+		this.validate(point);
+		this.root = this.insertNode(this.root, point, 0);
+		this._size++;
+	}
+
+	// ── KNN search ─────────────────────────────────────────────────────────────
+
+	/**
+	 * Find the k nearest neighbors to `query`.
+	 * Returns results sorted by distance ascending.
+	 */
+	knn(query: number[], k: number): KNNResult<T>[] {
+		if (k <= 0) throw new RangeError("k must be a positive integer");
+		this.validateVector(query);
+
+		const heap = new BoundedMaxHeap<T>(k);
+		this.searchKNN(this.root, query, k, heap, 0);
+		return heap.toSortedArray();
+	}
+
+	/**
+	 * Nearest single neighbor. Convenience wrapper around knn(query, 1).
+	 * Returns null if the tree is empty.
+	 */
+	nearest(query: number[]): KNNResult<T> | null {
+		const results = this.knn(query, 1);
+		return results[0] ?? null;
+	}
+
+	// ── Radius search ──────────────────────────────────────────────────────────
+
+	/**
+	 * Return all points whose distance to `query` is ≤ `radius`,
+	 * sorted by distance ascending.
+	 */
+	radiusSearch(query: number[], radius: number): KNNResult<T>[] {
+		if (radius < 0) throw new RangeError("radius must be non-negative");
+		this.validateVector(query);
+
+		const results: KNNResult<T>[] = [];
+		this.searchRadius(this.root, query, radius, results, 0);
+		results.sort((a, b) => a.distance - b.distance);
+		return results;
+	}
+
+	// ── Conversion ─────────────────────────────────────────────────────────────
+
+	/** Collect all points in the tree (order not guaranteed). */
+	toArray(): KDPoint<T>[] {
+		const out: KDPoint<T>[] = [];
+		this.collect(this.root, out);
+		return out;
+	}
+
+	/**
+	 * Rebuild the tree from its current points as a balanced tree.
+	 * Useful after many individual insertions to restore O(log n) query time.
+	 */
+	rebalance(): void {
+		const points = this.toArray();
+		this.root = points.length ? this.buildBalanced(points, 0) : null;
+	}
+
+	// ── Private: build ─────────────────────────────────────────────────────────
+
+	private buildBalanced(points: KDPoint<T>[], depth: number): KDNode<T> {
+		const axis = depth % this.dims;
+		points.sort((a, b) => a.vector[axis] - b.vector[axis]);
+
+		const mid = Math.floor(points.length / 2);
+		return {
+			point: points[mid],
+			axis,
+			left:  points.slice(0, mid).length
+				? this.buildBalanced(points.slice(0, mid), depth + 1)
+				: null,
+			right: points.slice(mid + 1).length
+				? this.buildBalanced(points.slice(mid + 1), depth + 1)
+				: null,
+		};
+	}
+
+	// ── Private: insert ────────────────────────────────────────────────────────
+
+	private insertNode(
+		node: KDNode<T> | null,
+		point: KDPoint<T>,
+		depth: number
+	): KDNode<T> {
+		if (node === null) {
+			return { point, axis: depth % this.dims, left: null, right: null };
+		}
+		const axis = depth % this.dims;
+		if (point.vector[axis] < node.point.vector[axis]) {
+			node.left = this.insertNode(node.left, point, depth + 1);
+		} else {
+			node.right = this.insertNode(node.right, point, depth + 1);
+		}
+		return node;
+	}
+
+	// ── Private: KNN traversal ─────────────────────────────────────────────────
+
+	private searchKNN(
+		node: KDNode<T> | null,
+		query: number[],
+		k: number,
+		heap: BoundedMaxHeap<T>,
+		depth: number
+	): void {
+		if (node === null) return;
+
+		const dist = this.distanceFn(query, node.point.vector);
+		heap.push({ point: node.point, distance: dist });
+
+		const axis = node.axis;
+		const diff = query[axis] - node.point.vector[axis];
+		const [near, far] = diff <= 0
+			? [node.left, node.right]
+			: [node.right, node.left];
+
+		this.searchKNN(near, query, k, heap, depth + 1);
+
+		// Only explore the far side if it could contain a closer point.
+		// For cosine distance we can't prune by axis gap alone, so always explore.
+		const shouldExplore =
+			this.distanceFn === cosine
+				? true
+				: Math.abs(diff) < heap.worstDistance;
+
+		if (shouldExplore) {
+			this.searchKNN(far, query, k, heap, depth + 1);
+		}
+	}
+
+	// ── Private: radius traversal ──────────────────────────────────────────────
+
+	private searchRadius(
+		node: KDNode<T> | null,
+		query: number[],
+		radius: number,
+		results: KNNResult<T>[],
+		depth: number
+	): void {
+		if (node === null) return;
+
+		const dist = this.distanceFn(query, node.point.vector);
+		if (dist <= radius) {
+			results.push({ point: node.point, distance: dist });
+		}
+
+		const axis = node.axis;
+		const diff = query[axis] - node.point.vector[axis];
+		const [near, far] = diff <= 0
+			? [node.left, node.right]
+			: [node.right, node.left];
+
+		this.searchRadius(near, query, radius, results, depth + 1);
+
+		const shouldExplore =
+			this.distanceFn === cosine ? true : Math.abs(diff) <= radius;
+
+		if (shouldExplore) {
+			this.searchRadius(far, query, radius, results, depth + 1);
+		}
+	}
+
+	// ── Private: collect ───────────────────────────────────────────────────────
+
+	private collect(node: KDNode<T> | null, out: KDPoint<T>[]): void {
+		if (node === null) return;
+		out.push(node.point);
+		this.collect(node.left, out);
+		this.collect(node.right, out);
+	}
+
+	// ── Private: validation ────────────────────────────────────────────────────
+
+	private validateVector(v: number[]): void {
+		if (v.length !== this.dims) {
+			throw new TypeError(
+				`Vector length ${v.length} does not match tree dimensionality ${this.dims}`
+			);
+		}
+	}
+
+	private validate(point: KDPoint<T>): void {
+		this.validateVector(point.vector);
+	}
+
+	private validateAll(points: KDPoint<T>[]): void {
+		for (const p of points) this.validate(p);
+	}
+}

src/llm.ts

@@ -1,16 +1,39 @@
-import {JSONAttemptParse} from '@ztimson/utils';
+import {snakeCase} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {Anthropic} from './antrhopic.ts';
-import {Ollama} from './ollama.ts';
 import {OpenAi} from './open-ai.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
-import {AiTool} from './tools.ts';
+import {AiTool, AiToolArg} from './tools.ts';
+import {fileURLToPath} from 'url';
+import {dirname, join} from 'path';
+import {spawn} from 'node:child_process';
+import {Memory, MemoryCache, MemoryManager, MemoryOptions} from './memory.ts';
+
+const MAX_AGENT_DEPTH = 5;
+
+export type AnthropicConfig = {proto: 'anthropic', token: string | string[]};
+export type OpenAiConfig = {proto: 'openai', host?: string, token: string | string[]};
+
+export type Agent = {
+	name: string;
+	description?: string;
+	model?: string | null;
+	temperature?: number;
+	system: string;
+	delegate?: boolean;
+	skills?: Skill[] | null;
+	tools?: AiTool[] | null;
+	mcp?: McpServer[] | null;
+	agents?: string[] | null;
+}
 
 export type LLMMessage = {
 	/** Message originator */
 	role: 'assistant' | 'system' | 'user';
 	/** Message content */
 	content: string | any;
+	/** Timestamp */
+	timestamp?: number;
 } | {
 	/** Tool call */
 	role: 'tool';
@@ -23,36 +46,18 @@ export type LLMMessage = {
 	/** Tool result */
 	content: undefined | string;
 	/** Tool error */
-	error: undefined | string;
+	error?: undefined | string;
+	/** Timestamp */
+	timestamp?: number;
+	/** Response duration in ms */
+	duration?: number;
+	/** Tokens per second */
+	tps?: number;
 }
 
-export type LLMOptions = {
-	/** Anthropic settings */
-	anthropic?: {
-		/** API Token */
-		token: string;
-		/** Default model */
-		model: string;
-	},
-	/** Ollama settings */
-	ollama?: {
-		/** connection URL */
-		host: string;
-		/** Default model */
-		model: string;
-	},
-	/** Open AI settings */
-	openAi?: {
-		/** API Token */
-		token: string;
-		/** Default model */
-		model: string;
-	},
-	/** Default provider & model */
-	model: string | [string, string];
-} & Omit<LLMRequest, 'model'>;
-
 export type LLMRequest = {
+	/** Return a parsed JSON object that matches the schema */
+	schema?: AiToolArg;
 	/** System prompt */
 	system?: string;
 	/** Message history */
@@ -64,56 +69,318 @@ export type LLMRequest = {
 	/** Available tools */
 	tools?: AiTool[];
 	/** LLM  model */
-	model?: string | [string, string];
+	model?: string;
 	/** Stream response */
-	stream?: (chunk: {text?: string, done?: true}) => any;
+	stream?: (chunk: {text?: string, tool?: string, done?: true}) => any;
 	/** Compress old messages in the chat to free up context */
-	compress?: {
+	compress?: {max: number; min: number};
-		/** Trigger chat compression once context exceeds the token count */
+	/** User's memory documents - RAG injected automatically each turn */
-		max: number;
+	memory?: Memory[] | MemoryCache | MemoryOptions;
-		/** Compress chat until context size smaller than */
+	/** Model to use for memory operations */
-		min: number
+	memoryModel?: string;
-	}
+	/** Skill documents the AI can browse and read on demand */
+	skills?: Skill[];
+	/** MCP servers to connect and expose as tools */
+	mcp?: McpServer[];
+	/** Subagents exposed as delegatable/wrapped tools */
+	agents?: Agent[];
+	/** @internal recursion guard for nested agent delegation */
+	_agentDepth?: number;
 }
 
-export class LLM {
+export type McpServer = {
-	private providers: {[key: string]: LLMProvider} = {};
+	/** MCP server name for humans */
+	name: string;
+	/** Host URL */
+	host: string;
+	/** Server access token */
+	token?: string;
+}
 
-	constructor(public readonly ai: Ai, public readonly options: LLMOptions) {
+export type Skill = {
-		if(options.anthropic?.token) this.providers.anthropic = new Anthropic(this.ai, options.anthropic.token, options.anthropic.model);
+	/** Name of skill for humans */
-		if(options.ollama?.host) this.providers.ollama = new Ollama(this.ai, options.ollama.host, options.ollama.model);
+	name: string;
-		if(options.openAi?.token) this.providers.openAi = new OpenAi(this.ai, options.openAi.token, options.openAi.model);
+	/** Description LLM will use to decide to learn a skill */
+	description: string;
+	/** Skill instructions */
+	content: string;
+}
+
+class LLM {
+	private memoryManager!: MemoryManager;
+
+	defaultModel!: string;
+	models: {[model: string]: LLMProvider} = {};
+
+	constructor(public readonly ai: Ai) {
+		if(!ai.options.llm?.models) return;
+		Object.entries(ai.options.llm.models).forEach(([model, config]) => {
+			if(!this.defaultModel) this.defaultModel = model;
+			if(config.proto == 'anthropic') this.models[model] = new Anthropic(this.ai, config.token, model);
+			else if(config.proto == 'openai') this.models[model] = new OpenAi(this.ai, config.host || null, config.token, model);
+		});
+		this.memoryManager = new MemoryManager(this);
+	}
+
+	private setupAgent(agents: Agent[] = [], allAgents: Agent[], history: LLMMessage[], aborts: (() => void)[], depth = 0, delegateState: {resp: string | null}): AiTool[] {
+		return agents.map(a => {
+			const toolName = `${a.delegate ? '' : 'sub'}agent_${snakeCase(a.name)}`;
+			return {
+				name: toolName,
+				description: `${a.delegate ? 'Delegate to ' : ''}Subagent: ${a.description || a.name}`,
+				args: <any>(a.delegate ? {} : {
+					context: {type: 'string', description: 'Summary of related messages, samples, files, etc...', required: true},
+					instructions: {type: 'string', description: 'Detailed instructions for subagent to complete', required: true},
+				}),
+				fn: async (args: any, stream: any, ai: any, id?: string) => {
+					if(depth >= MAX_AGENT_DEPTH) return 'Max agent delegation depth exceeded';
+
+					// Opt-in only, self always excluded regardless of whitelist
+					const nested = (a.agents || [])
+						.map(name => allAgents.find(x => x.name === name))
+						.filter((x): x is Agent => !!x && x.name !== a.name);
+
+					const request = this.ask(a.delegate ? '' : `${args.instructions}${args.context ? `\n\n<context>${args.context}</context>` : ''}`, {
+						system: `You are a specialized subagent. ${a.delegate ? 'Your output streams directly to the user for the remainder of this turn. You are mid conversation - dispense with greetings.' : 'You are wrapped in a tool call that will be analysis by an LLM - dispense with conversation'}
+As a subagent, focus on executing your task completely using available tools and returning only the final result - no commentary, questions, or dialogue.
+
+${a.system}`,
+						model: a.model || undefined,
+						temperature: a.temperature,
+						stream: a.delegate ? stream : undefined,
+						history: a.delegate ? history : [],
+						mcp: a.mcp || undefined,
+						skills: a.skills || undefined,
+						tools: a.tools || undefined,
+						agents: nested,
+						_agentDepth: depth + 1,
+					} as any);
+					aborts.push(request.abort);
+					const resp = await request;
+
+					if(a.delegate) {
+						delegateState.resp = resp;
+						return '';
+					}
+					return resp;
+				}
+			};
+		});
+	}
+
+	private async setupMcp(servers: McpServer[] = []): Promise<{prompt: string, tools: AiTool[]}> {
+		if(!servers?.length) return {prompt: '', tools: []};
+		const allTools: AiTool[] = [];
+		await Promise.all(servers.map(async server => {
+			const res = await fetch(`${server.host}/tools`, {headers: server.token ? {Authorization: `Bearer ${server.token}`} : {}});
+			const mcp: any = await res.json();
+			if(!mcp?.tools) return;
+			for(const t of mcp.tools) {
+				const args: Record<string, any> = {};
+				if(t.inputSchema?.properties) {
+					for(const [key, val] of Object.entries<any>(t.inputSchema.properties)) {
+						args[key] = {type: val.type || 'string', description: val.description || '', required: t.inputSchema.required?.includes(key)};
+					}
+				}
+				allTools.push({
+					name: `${server.name}_${t.name}`,
+					description: t.description || '',
+					args,
+					fn: async (a: any) => {
+						const r = await fetch(`${server.host}/tools/call`, {
+							method: 'POST',
+							headers: {'Content-Type': 'application/json', ...(server.token ? {Authorization: `Bearer ${server.token}`} : {})},
+							body: JSON.stringify({name: t.name, arguments: a})
+						});
+						const data: any = await r.json();
+						return data?.content?.[0]?.text ?? JSON.stringify(data);
+					}
+				});
+			}
+		}));
+
+		const list = allTools.map(t => `- ${t.name}: ${t.description}`).join('\n');
+		return {
+			prompt: `You have access to the following MCP tools:\n${list}`,
+			tools: allTools
+		};
+	}
+
+	private setupSkills(skills: Skill[] = []): {prompt: string, tools: AiTool[]} {
+		if(!skills?.length) return {prompt: '', tools: []};
+		const list = skills.map(s => `- ${s.name}: ${s.description}`).join('\n');
+		return {
+			prompt: `You have access to the following skill documents, use \`read_skill\` to access them:\n${list}`,
+			tools: [{
+				name: 'skill_read',
+				description: 'Read the full content of a skill/knowledge document',
+				args: {
+					name: {type: 'string', description: 'Exact skill name', required: true}
+				},
+				fn: (args: any) => {
+					const skill = skills.find(s => s.name === args.name);
+					if(!skill) return `Skill not found. Available:\n${list}`;
+					return `# ${skill.name}\n${skill.content}`;
+				}
+			}]
+		}
+	}
+
+	private wrapToolTiming(tools: AiTool[], timings: Map<string, {duration: number, tps: number}>): AiTool[] {
+		return tools.map(t => ({
+			...t,
+			fn: async (args: any, stream: any, ai: any, id?: string) => {
+				const start = Date.now();
+				const result = await t.fn(args, stream, ai, id);
+				const duration = Date.now() - start;
+				const tps = duration > 0 ? this.estimateTokens(result) / (duration / 1000) : 0;
+				if(id) timings.set(id, {duration, tps});
+				return result;
+			}
+		}));
+	}
+
+	ask(message: string, options: LLMRequest = {}): AbortablePromise<string> {
+		options = <any>{
+			system: '',
+			...this.ai.options.llm,
+			models: undefined,
+			history: [],
+			...options,
+		}
+		const m = options.model || this.defaultModel;
+		if(!this.models[m]) throw new Error(`Model does not exist: ${m}`);
+		let request: AbortablePromise<string> | null = null;
+		let aborted = false;
+		const nestedAborts: (() => void)[] = [];
+		const abort = () => {
+			aborted = true;
+			request?.abort?.();
+			nestedAborts.forEach(a => a());
+		};
+
+		let promise: any;
+		const requestStart = Date.now();
+
+		promise = (async () => {
+			let tools: AiTool[] = options.tools || this.ai.options.llm?.tools || [];
+			const prompts: string[] = [];
+			let history = options.history || [];
+
+			// MCP
+			const mcp = options.mcp || this.ai.options?.llm?.mcp;
+			if(mcp?.length) {
+				const m = await this.setupMcp(mcp);
+				prompts.unshift(m.prompt);
+				tools.push(...m.tools);
+			}
+
+			// Skills
+			const skills = options.skills || this.ai.options?.llm?.skills;
+			if(skills?.length) {
+				const s = this.setupSkills(skills);
+				prompts.unshift(s.prompt);
+				tools.push(...s.tools);
+			}
+
+			// Agents
+			const agents = options.agents || this.ai.options?.llm?.agents;
+			const delegateState: {resp: string | null} = {resp: null};
+			if(agents?.length) tools.push(...this.setupAgent(agents, agents, history, nestedAborts, options._agentDepth || 0, delegateState));
+
+			// Memory
+			const mem = MemoryManager.normalize(options.memory);
+			if(mem) {
+				const mems = mem.memory instanceof MemoryCache ? mem.memory.memories : mem.memory;
+				if(mems.length) {
+					if(mem.inject) {
+						const pool = 15; // candidates considered, cheap since only refs are listed
+						const budget = mem.maxTokens ?? 2000; // actual content injected
+						const relevant = await this.memoryManager.recollect(message, mem.memory, pool);
+
+						let used = 0;
+						const preloaded: typeof relevant = [];
+						const listed: typeof relevant = [];
+						for(const r of relevant) {
+							const t = this.estimateTokens(r.content);
+							if(used + t <= budget || preloaded.length === 0) {
+								preloaded.push(r);
+								used += t;
+							} else listed.push(r);
+						}
+
+						prompts.unshift(`You have access to the following memory files:
+${mems.map(m => `- ${m.name}: ${m.description}`).join('\n')}
+${preloaded.length ? `
+Relevant memories have been preloaded:
+${preloaded.map(r => `
+**${r.name}**
+${r.description}
+${r.content}
+`).join('\n---\n')}
+` : ''}${listed.length ? `
+Also relevant but not preloaded (use \`memory_recall\`): ${listed.map(r => r.name).join(', ')}
+` : ''}`.trim());
+					}
+					if(mem.tool) tools.push(this.memoryManager.tools.read(mem.memory));
+				}
+			}
+
+			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
+
+			const toolTimings = new Map<string, {duration: number, tps: number}>();
+			tools = this.wrapToolTiming(tools, toolTimings);
+
+			if(aborted) throw Object.assign(new Error('Aborted'), {name: 'AbortError'});
+
+			prompts.unshift(options.system || this.ai.options.llm?.system || '');
+			request = this.models[m].ask(message, {...options, tools, system: prompts.filter(Boolean).join('\n\n')});
+			let resp = await request;
+
+			// Capture meta (duration / tps)
+			for(const h of history) {
+				if(h.role === 'tool' && toolTimings.has(h.id)) Object.assign(h, toolTimings.get(h.id));
+			}
+
+			if(typeof resp === 'string' && !resp.trim() && delegateState.resp !== null) resp = delegateState.resp;
+
+			if(mem?.tool) history.splice(0, history.length, ...history.filter(h => h.role !== 'tool' || h.name !== 'memory_recall'));
+			if(options.compress && this.estimateTokens(history) >= options.compress.max) {
+				if(mem?.update) await this.memoryManager.memorize(history, mem.memory, {model: options.memoryModel || this.defaultModel, ...options});
+				const compressed = await this.compressHistory(history, options.compress.max, options.compress.min, options);
+				if(options.history) options.history.splice(0, options.history.length, ...compressed);
+			}
+
+			const requestDuration = Date.now() - requestStart;
+			const totalTokens = history
+				.filter((h: any) => h.role === 'assistant' && h.duration && h.tps)
+				.reduce((sum: number, h: any) => sum + h.tps * (h.duration / 1000), 0);
+			const requestTps = requestDuration > 0 ? totalTokens / (requestDuration / 1000) : 0;
+			Object.assign(promise, {duration: requestDuration, tps: requestTps});
+
+			return resp;
+		})();
+
+		return Object.assign(promise, {abort});
 	}
 
 	/**
-	 * Chat with LLM
+	 * Digest full conversation history into memory documents.
-	 * @param {string} message Question
+	 * Call on session end to persist the conversation.
-	 * @param {LLMRequest} options Configuration options and chat history
-	 * @returns {{abort: () => void, response: Promise<LLMMessage[]>}} Function to abort response and chat history
 	 */
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
+	async updateMemory(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest = {}): Promise<Memory[]> {
-		let model: any = [null, null];
+		return this.memoryManager.memorize(history, memories, {model: this.defaultModel, ...options});
-		if(options.model) {
-			if(typeof options.model == 'object') model = options.model;
-			else model = [options.model, (<any>this.options)[options.model]?.model];
-		}
-		if(!options.model || model[1] == null) {
-			if(typeof this.options.model == 'object') model = this.options.model;
-			else model = [this.options.model, (<any>this.options)[this.options.model]?.model];
-		}
-		if(!model[0] || !model[1]) throw new Error(`Unknown LLM provider or model: ${model[0]} / ${model[1]}`);
-		return this.providers[model[0]].ask(message, {...options, model: model[1]});
 	}
 
 	/**
 	 * Compress chat history to reduce context size
 	 * @param {LLMMessage[]} history Chatlog that will be compressed
 	 * @param max Trigger compression once context is larger than max
-	 * @param min Summarize until context size is less than min
+	 * @param min Leave messages less than the token minimum, summarize the rest
 	 * @param {LLMRequest} options LLM options
 	 * @returns {Promise<LLMMessage[]>} New chat history will summary at index 0
 	 */
-	async compress(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<LLMMessage[]> {
+	async compressHistory(history: LLMMessage[], max: number, min: number, options?: LLMRequest): Promise<LLMMessage[]> {
 		if(this.estimateTokens(history) < max) return history;
 		let keep = 0, tokens = 0;
 		for(let m of history.toReversed()) {
@@ -122,10 +389,121 @@ export class LLM {
 			else break;
 		}
 		if(history.length <= keep) return history;
-		const recent = keep == 0 ? [] : history.slice(-keep),
+		const system = history[0].role == 'system' ? history[0] : null,
+			recent = keep == 0 ? [] : history.slice(-keep),
 			process = (keep == 0 ? history : history.slice(0, -keep)).filter(h => h.role === 'assistant' || h.role === 'user');
-		const summary = await this.summarize(process.map(m => `${m.role}: ${m.content}`).join('\n\n'), 250, options);
+
-		return [{role: 'assistant', content: `Conversation Summary: ${summary}`}, ...recent];
+		const summary: any = await this.summarize(process.map(m => `[${m.role}]: ${m.content}`).join('\n\n'), 500, options);
+		const d = Date.now();
+		const h = [{role: <any>'tool', name: 'summary', id: `summary_` + d, args: {}, content: `Conversation Summary: ${summary?.summary}`, timestamp: d}, ...recent];
+		if(system) h.splice(0, 0, system);
+		return h;
+	}
+
+	/**
+	 * Compare the difference between embeddings (calculates the angle between two vectors)
+	 * @param {number[]} v1 First embedding / vector comparison
+	 * @param {number[]} v2 Second embedding / vector for comparison
+	 * @returns {number} Similarity values 0-1: 0 = unique, 1 = identical
+	 */
+	cosineSimilarity(v1: number[], v2: number[]): number {
+		if (v1.length !== v2.length) throw new Error('Vectors must be same length');
+		let dotProduct = 0, normA = 0, normB = 0;
+		for (let i = 0; i < v1.length; i++) {
+			dotProduct += v1[i] * v2[i];
+			normA += v1[i] * v1[i];
+			normB += v2[i] * v2[i];
+		}
+		const denominator = Math.sqrt(normA) * Math.sqrt(normB);
+		return denominator === 0 ? 0 : dotProduct / denominator;
+	}
+
+	/**
+	 * Chunk text into parts for AI digestion
+	 * @param {object | string} target Item that will be chunked (objects get converted)
+	 * @param {number} maxTokens Chunking size. More = better context, less = more specific (Search by paragraphs or lines)
+	 * @param {number} overlapTokens Includes previous X tokens to provide continuity to AI (In addition to max tokens)
+	 * @returns {string[]} Chunked strings
+	 */
+	chunk(target: object | string, maxTokens = 500, overlapTokens = 50): string[] {
+		const objString = (obj: any, path = ''): string[] => {
+			if(!obj) return [];
+			return Object.entries(obj).flatMap(([key, value]) => {
+				const p = path ? `${path}${isNaN(+key) ? `.${key}` : `[${key}]`}` : key;
+				if(typeof value === 'object' && !Array.isArray(value)) return objString(value, p);
+				return `${p}: ${Array.isArray(value) ? value.join(', ') : value}`;
+			});
+		};
+		const lines = typeof target === 'object' ? objString(target) : target.toString().split('\n');
+		const tokens = lines.flatMap(l => [...l.split(/\s+/).filter(Boolean), '\n']);
+		const chunks: string[] = [];
+		for(let i = 0; i < tokens.length;) {
+			let text = '', j = i;
+			while(j < tokens.length) {
+				const next = text + (text ? ' ' : '') + tokens[j];
+				if(this.estimateTokens(next.replace(/\s*\n\s*/g, '\n')) > maxTokens && text) break;
+				text = next;
+				j++;
+			}
+			const clean = text.replace(/\s*\n\s*/g, '\n').trim();
+			if(clean) chunks.push(clean);
+			i = Math.max(j - overlapTokens, j === i ? i + 1 : j);
+		}
+		return chunks;
+	}
+
+	/**
+	 * Create a vector representation of a string
+	 * @param {object | string} target Item that will be embedded (objects get converted)
+	 * @param {maxTokens?: number, overlapTokens?: number} opts Options for embedding such as chunk sizes
+	 * @returns {Promise<Awaited<{index: number, embedding: number[], text: string, tokens: number}>[]>} Chunked embeddings
+	 */
+	embedding(target: object | string, opts: {maxTokens?: number, overlapTokens?: number} = {}): AbortablePromise<{index: number, embedding: number[], text: string, tokens: number}[]> {
+		let {maxTokens = 500, overlapTokens = 50} = opts;
+		let aborted = false;
+		const abort = () => { aborted = true; };
+
+		const embed = (text: string): Promise<number[]> => {
+			return new Promise((resolve, reject) => {
+				if(aborted) return reject(new Error('Aborted'));
+				const args: string[] = [
+					join(dirname(fileURLToPath(import.meta.url)), 'embedder.js'),
+					<string>this.ai.options.path,
+					this.ai.options?.embedder || 'bge-small-en-v1.5'
+				];
+				const proc = spawn('node', args, {stdio: ['pipe', 'pipe', 'ignore']});
+				proc.stdin.write(text);
+				proc.stdin.end();
+				let output = '';
+				proc.stdout.on('data', (data: Buffer) => output += data.toString());
+				proc.on('close', (code: number) => {
+					if(aborted) return reject(new Error('Aborted'));
+					if(code === 0) {
+						try {
+							const result = JSON.parse(output);
+							resolve(result.embedding);
+						} catch(err) {
+							reject(err);
+						}
+					} else {
+						reject(new Error(`Embedder process exited with code ${code}`));
+					}
+				});
+				proc.on('error', reject);
+			});
+		};
+
+		const p = (async () => {
+			const chunks = this.chunk(target, maxTokens, overlapTokens), results: any[] = [];
+			for(let i = 0; i < chunks.length; i++) {
+				if(aborted) break;
+				const text = chunks[i];
+				const embedding = await embed(text);
+				results.push({index: i, embedding, text, tokens: this.estimateTokens(text)});
+			}
+			return results;
+		})();
+		return <any>Object.assign(p, {abort});
 	}
 
 	/**
@@ -139,29 +517,96 @@ export class LLM {
 	}
 
 	/**
-	 * Ask a question with JSON response
+	 * Compare the difference between two strings using tensor math
-	 * @param {string} message Question
+	 * @param target Text that will be checked
-	 * @param {LLMRequest} options Configuration options and chat history
+	 * @param {string} searchTerms Multiple search terms to check against target
-	 * @returns {Promise<{} | {} | RegExpExecArray | null>}
+	 * @returns {{avg: number, max: number, similarities: number[]}} Similarity values 0-1: 0 = unique, 1 = identical
 	 */
-	async json(message: string, options?: LLMRequest) {
+	fuzzyMatch(target, ...searchTerms) {
-		let resp = await this.ask(message, {
+		if (searchTerms.length < 2) throw new Error('Requires at least 2 strings to compare');
-			system: 'Respond using a JSON blob',
+		const levenshtein = (a, b) => {
-			...options
+			const m = a.length, n = b.length;
-		});
+			if (!m) return n;
-		if(!resp?.[0]?.content) return {};
+			if (!n) return m;
-		return JSONAttemptParse(new RegExp('\{[\s\S]*\}').exec(resp[0].content), {});
+			const dp = Array.from({length: m + 1}, (_, i) => [i, ...Array(n).fill(0)]);
+			for (let j = 0; j <= n; j++) dp[0][j] = j;
+			for (let i = 1; i <= m; i++) {
+				for (let j = 1; j <= n; j++) {
+					dp[i][j] = a[i - 1] === b[j - 1]
+						? dp[i - 1][j - 1]
+						: 1 + Math.min(dp[i - 1][j - 1], dp[i - 1][j], dp[i][j - 1]);
+				}
+			}
+			return dp[m][n];
+		};
+		const similarity = (a, b) => {
+			a = a.toLowerCase(); b = b.toLowerCase();
+			return 1 - levenshtein(a, b) / Math.max(a.length, b.length, 1);
+		};
+		const similarities = searchTerms.map(t => similarity(target, t));
+		return {
+			avg: similarities.reduce((acc, s) => acc + s, 0) / similarities.length,
+			max: Math.max(...similarities),
+			similarities
+		};
 	}
 
 	/**
 	 * Create a summary of some text
 	 * @param {string} text Text to summarize
-	 * @param {number} tokens Max number of tokens
+	 * @param {number} length Max number of words
 	 * @param options LLM request options
 	 * @returns {Promise<string>} Summary
 	 */
-	summarize(text: string, tokens: number, options?: LLMRequest): Promise<string | null> {
+	async summarize(text: string, length: number = 500, options?: LLMRequest): Promise<string | null> {
-		return this.ask(text, {system: `Generate a brief summary <= ${tokens} tokens. Output nothing else`, temperature: 0.3, ...options})
+		let system = `Your job is to summarize the users message using tool calls. Call the \`submit\` tool at least once with the shortest summary possible that's <= ${length} words. The tool call will respond with the token count. Responses are ignored`;
-			.then(history => <string>history.pop()?.content || null);
+		if(options?.system) system += '\n\n' + options.system;
+		return new Promise(async (resolve, reject) => {
+			let done = false;
+			const resp = await this.ask(text, {
+				temperature: 0.3,
+				...options,
+				system,
+				tools: [{
+					name: 'submit',
+					description: 'Submit summary',
+					args: {summary: {type: 'string', description: 'Text summarization', required: true}},
+					fn: (args) => {
+						if(!args.summary) return 'No summary provided';
+						const count = args.summary.split(' ').length;
+						if(count > length) return `Too long: ${length} words`;
+						done = true;
+						resolve(args.summary || null);
+						return `Saved: ${length} words`;
+					}
+				}, ...(options?.tools || [])],
+			});
+			if(!done) reject(`AI failed to create summary:\n${resp}`);
+		});
+	}
+
+	addModel(name: string, config: AnthropicConfig | OpenAiConfig, setDefault = false) {
+		if(config.proto == 'anthropic') this.models[name] = new Anthropic(this.ai, config.token, name);
+		else if(config.proto == 'openai') this.models[name] = new OpenAi(this.ai, config.host || null, config.token, name);
+		if(setDefault || !this.defaultModel) this.defaultModel = name;
+	}
+
+	removeModel(name: string) {
+		delete this.models[name];
+		if(this.defaultModel === name) {
+			this.defaultModel = Object.keys(this.models)[0] ?? '';
+		}
+	}
+
+	setModels(models: {[model: string]: AnthropicConfig | OpenAiConfig}, replace = true) {
+		if(replace) this.models = {};
+		Object.entries(models).forEach(([model, config]) => {
+			if(!this.defaultModel) this.defaultModel = model;
+			if(config.proto == 'anthropic') this.models[model] = new Anthropic(this.ai, config.token, model);
+			else if(config.proto == 'openai') this.models[model] = new OpenAi(this.ai, config.host || null, config.token, model);
+		});
+		this.defaultModel = Object.keys(this.models)[0] ?? '';
 	}
 }
+
+export default LLM;

src/memory.ts

@@ -0,0 +1,502 @@
+import {LLMRequest, LLMMessage} from './llm.ts';
+import {AiTool} from './tools.ts';
+import {KDPoint, KDTree} from './kd-tree.ts';
+
+const FACTS_HEADING = '## Facts';
+
+const GENERIC_TEMPLATE = `# {{Title}}
+
+## Summary
+
+## Details
+
+## Related`;
+
+export class MemoryCache {
+	private tree: KDTree<MemoryRef>;
+	public memories: Memory[];
+
+	get length() { return this.memories.length; }
+
+	constructor(memories: Memory[]) {
+		this.memories = memories;
+		this.tree = this.buildTree();
+	}
+
+	private buildTree(): KDTree<MemoryRef> {
+		const embedded = this.memories.filter(m => m.embedding?.length);
+		if (!embedded.length) return new KDTree<MemoryRef>(0);
+
+		const dims = embedded[0].embedding.length;
+		const points: KDPoint<MemoryRef>[] = embedded.map(m => ({
+			vector: m.embedding,
+			payload: {name: m.name, description: m.description},
+		}));
+
+		return new KDTree<MemoryRef>(dims, 'cosine', points);
+	}
+
+	search(query: number[], limit: number): MemoryRef[] {
+		const results = this.tree.knn(query, limit);
+		return results.map(r => r.point.payload);
+	}
+
+	add(memory: Memory): void {
+		this.memories.push(memory);
+		this.rebuild();
+	}
+
+	update(memory: Memory): void {
+		const idx = this.memories.findIndex(m => m.name === memory.name);
+		if (idx !== -1) {
+			this.memories[idx] = memory;
+		} else {
+			this.memories.push(memory);
+		}
+		this.rebuild();
+	}
+
+	remove(name: string): void {
+		const idx = this.memories.findIndex(m => m.name === name);
+		if (idx !== -1) {
+			this.memories.splice(idx, 1);
+			this.rebuild();
+		}
+	}
+
+	rebuild(): void {
+		this.tree = this.buildTree();
+	}
+}
+
+export type MemoryOptions = {
+	/** Memory object */
+	memory: Memory[] | MemoryCache;
+	/** Inject N memories into the system prompt */
+	inject?: boolean;
+	/** expose recall tool to LLM */
+	tool?: boolean;
+	/** Update memory on compression */
+	update?: boolean;
+	/** Max context size of memories to inject to each call (removed immediately after use) */
+	maxTokens?: number;
+}
+
+export type Memory = {
+	name: string;
+	description: string;
+	content: string;
+	embedding: number[];
+	links: string[];
+	backlinks: string[];
+}
+
+type MemoryRef = {
+	name: string;
+	description: string;
+}
+
+type FactBucket = {
+	subject: string;
+	facts: string[];
+}
+
+function extractLinks(content: string): string[] {
+	if (!content) return [];
+	const matches = content.matchAll(/\[\[([^\]]+)\]\]/g);
+	return [...new Set([...matches].map(m => m[1].trim()))];
+}
+
+export function rebuildGraph(memories: Memory[]): void {
+	for (const m of memories) m.links = extractLinks(m.content).filter(l => l !== m.name);
+	for (const m of memories) m.backlinks = [];
+	for (const m of memories) {
+		for (const link of m.links) {
+			const target = memories.find(t => t.name === link);
+			if (target) target.backlinks.push(m.name);
+		}
+	}
+}
+
+function dedupeFacts(facts: string[]): string[] {
+	const seen = new Map<string, string>();
+	for (const f of facts) {
+		const clean = f.trim();
+		if (clean) seen.set(clean.toLowerCase(), clean);
+	}
+	return [...seen.values()];
+}
+
+function cosineDistance(a: number[], b: number[]): number {
+	let dot = 0, normA = 0, normB = 0;
+	for (let i = 0; i < a.length; i++) {
+		dot   += a[i] * b[i];
+		normA += a[i] * a[i];
+		normB += b[i] * b[i];
+	}
+	const denom = Math.sqrt(normA) * Math.sqrt(normB);
+	return denom === 0 ? 1 : 1 - dot / denom;
+}
+
+function getWeekMonday(date: Date = new Date()): string {
+	const d = new Date(Date.UTC(date.getFullYear(), date.getMonth(), date.getDate()));
+	const day = d.getUTCDay();
+	const diff = day === 0 ? -6 : 1 - day;
+	d.setUTCDate(d.getUTCDate() + diff);
+	return d.toISOString().slice(0, 10);
+}
+
+export class MemoryManager {
+	private recentlyTouched = new Map<string, number>();
+
+	private queues = new Map<string, {
+		dirty: boolean,
+		request: {abort?: () => void} | null,
+		task: Promise<void>,
+	}>();
+
+	tools = {
+		read: (memories: Memory[] | MemoryCache): AiTool => ({
+			name: 'memory_recall',
+			description: 'Read the full content of a memory document',
+			args: {
+				name: {type: 'string', description: 'Exact memory name', required: true},
+			},
+			fn: (args: any) => {
+				const mems = this.unwrap(memories);
+				const mem = mems.find(m => m.name === args.name);
+				if (!mem) return 'Document not found';
+				this.touch(mem.name);
+				return mem.content;
+			},
+		}),
+
+		forget: (memories: Memory[] | MemoryCache): AiTool => ({
+			name: 'memory_forget',
+			description: 'Permanently delete a memory document and clean up all references to it',
+			args: {
+				name: {type: 'string', description: 'Exact memory name to forget', required: true}
+			},
+			fn: (args: any) => {
+				const result = this.forget(args.name, memories);
+				return result ? `Forgotten: ${args.name}` : `Not found: ${args.name}`;
+			},
+		}),
+	};
+
+	constructor(private llm: any) {}
+
+	static normalize(m?: Memory[] | MemoryCache | MemoryOptions) {
+		if(!m) return null;
+		const raw = m instanceof MemoryCache || Array.isArray(m);
+		return raw ? {memory: <Memory[] | MemoryCache>m, inject: true, tool: true, update: true} : {inject: true, tool: true, update: true, ...m};
+	}
+
+	private unwrap(memories: Memory[] | MemoryCache): Memory[] {
+		return memories instanceof MemoryCache ? memories.memories : memories;
+	}
+
+	private sync(memories: Memory[] | MemoryCache): void {
+		if (memories instanceof MemoryCache) memories.rebuild();
+	}
+
+	private parseFrontmatter(content: string): {fm: Map<string, string>, body: string} {
+		const match = content.match(/^---\n([\s\S]*?)\n---\n?([\s\S]*)$/);
+		if (!match) return {fm: new Map(), body: content};
+		const fm = new Map<string, string>();
+		for (const line of match[1].split('\n')) {
+			const i = line.indexOf(':');
+			if (i === -1) continue;
+			fm.set(line.slice(0, i).trim(), line.slice(i + 1).trim());
+		}
+		return {fm, body: match[2]};
+	}
+
+	private writeFrontmatter(fm: Map<string, string>, body: string): string {
+		const lines = [...fm.entries()].map(([k, v]) => `${k}: ${v}`);
+		return `---\n${lines.join('\n')}\n---\n\n${body.trimStart()}`;
+	}
+
+	private stripHeader(content: string): string {
+		return content.replace(/^---[\s\S]*?\n---\n?/, '').trimStart();
+	}
+
+	private touchHeader(node: Memory, body: string): string {
+		const {fm} = this.parseFrontmatter(node.content);
+		fm.set('name', node.name);
+		fm.set('description', node.description || '');
+		fm.set('modified', new Date().toISOString());
+		return this.writeFrontmatter(fm, body);
+	}
+
+	private ensureDoc(node: Memory): void {
+		if (node.content) return;
+		const title = node.name.split('/').pop() ?? node.name;
+		node.content = this.touchHeader(node, `# ${title}\n`);
+	}
+
+	private appendFacts(node: Memory, facts: string[]): void {
+		this.ensureDoc(node);
+		const body = this.stripHeader(node.content);
+		const bullets = facts.map(f => `- ${f}`).join('\n');
+		const idx = body.indexOf(FACTS_HEADING);
+		const newBody = idx === -1
+			? `${body.trimEnd()}\n\n${FACTS_HEADING}\n${bullets}\n`
+			: `${body.slice(0, idx + FACTS_HEADING.length)}\n${bullets}${body.slice(idx + FACTS_HEADING.length)}`;
+		node.content = this.touchHeader(node, newBody);
+	}
+
+	decay() {
+		for(const [name, ttl] of this.recentlyTouched) {
+			if(ttl <= 1) this.recentlyTouched.delete(name);
+			else this.recentlyTouched.set(name, ttl - 1);
+		}
+	}
+
+	touch(name: string, ttl = 2) {
+		this.recentlyTouched.set(name, ttl);
+	}
+
+	getTouched(): string[] {
+		return [...this.recentlyTouched.keys()];
+	}
+
+	forget(name: string, memories: Memory[] | MemoryCache): boolean {
+		const mem = this.unwrap(memories);
+		const idx = mem.findIndex(m => m.name === name);
+		if (idx === -1) return false;
+
+		mem.splice(idx, 1);
+		rebuildGraph(mem);
+		this.sync(memories);
+		return true;
+	}
+
+	async recollect(query: string, memories: Memory[] | MemoryCache, limit = 5, graphDepth = 1): Promise<Memory[]> {
+		const mem = this.unwrap(memories);
+		if (!mem.length) return [];
+
+		const [e] = await this.llm.embedding(query);
+		if (!e) return [];
+
+		let vectorResults: MemoryRef[];
+		if (memories instanceof MemoryCache) vectorResults = memories.search(e.embedding, limit);
+		else vectorResults = this.cosineSearch(e.embedding, mem, limit);
+		const found = new Set<string>(vectorResults.map(r => r.name));
+
+		if (graphDepth > 0) {
+			const frontier = [...found];
+			for (let depth = 0; depth < graphDepth; depth++) {
+				const next: string[] = [];
+				for (const name of frontier) {
+					const node = mem.find(m => m.name === name);
+					if (!node) continue;
+					for (const link of node.links) {
+						if (!found.has(link) && mem.find(m => m.name === link)) {
+							found.add(link);
+							next.push(link);
+						}
+					}
+				}
+				frontier.splice(0, frontier.length, ...next);
+				if (!frontier.length) break;
+			}
+		}
+
+		const vectorOrder = vectorResults.map(r => r.name);
+		const graphExpansions = [...found].filter(n => !vectorOrder.includes(n));
+		const ordered = [...vectorOrder, ...graphExpansions];
+		return ordered.map(n => mem.find(m => m.name === n)!).filter(Boolean);
+	}
+
+	private cosineSearch(query: number[], memories: Memory[], limit: number): MemoryRef[] {
+		const scored = memories
+			.filter(m => m.embedding?.length)
+			.map(m => ({
+				ref: {name: m.name, description: m.description},
+				distance: cosineDistance(query, m.embedding),
+			}))
+			.sort((a, b) => a.distance - b.distance)
+			.slice(0, limit);
+		return scored.map(s => s.ref);
+	}
+
+	private listNodes(memories: Memory[]): MemoryRef[] {
+		return memories.map(m => ({name: m.name, description: m.description}));
+	}
+
+	async memorize(history: LLMMessage[], memories: Memory[] | MemoryCache, options: LLMRequest): Promise<Memory[]> {
+		const conversation = history
+			.filter(h => h.role === 'user' || h.role === 'assistant')
+			.map(h => `[${h.role}]: ${h.content}`).join('\n\n').trim();
+		if (!conversation) return [];
+
+		const uid = `${Date.now()}_${Math.random().toString(36).slice(2)}`;
+		// NOTE: adjust field names below (id/tool_call_id/name) to match your LLMMessage/tool-call schema.
+		const pending = {role: 'tool', name: 'memory_process', id: uid, content: 'Processing…'} as unknown as LLMMessage;
+		history.push(pending);
+
+		const mem = this.unwrap(memories);
+		const buckets = await this.factAgent(conversation, mem, options, getWeekMonday());
+		const touched: Memory[] = [];
+
+		for (const {subject, facts} of buckets) {
+			let node = mem.find(m => m.name === subject);
+			if (!node) {
+				node = {name: subject, description: '', content: '', embedding: [], links: [], backlinks: []};
+				mem.push(node);
+			}
+			this.appendFacts(node, facts);
+			const [e] = await this.llm.embedding(node.content);
+			if (e) node.embedding = e.embedding;
+			this.touch(node.name);
+			touched.push(node);
+		}
+
+		if (touched.length) {
+			rebuildGraph(mem);
+			this.sync(memories);
+			(pending as any).content = `Saved to ${touched.map(n => `[[${n.name}]]`).join(', ')}`;
+			for (const node of touched) this.reconcile(node, memories, options).catch(() => {});
+		} else {
+			(pending as any).content = 'Nothing worth remembering.';
+		}
+
+		return touched;
+	}
+
+	/** Manual/cron entry point. scope 'touched' only reconciles docs with a pending Facts inbox. */
+	async reconcileVault(memories: Memory[] | MemoryCache, options: LLMRequest, scope: 'touched' | 'all' = 'touched'): Promise<void> {
+		const mem = this.unwrap(memories);
+		const targets = scope === 'all' ? mem : mem.filter(m => m.content.includes(FACTS_HEADING));
+		await Promise.all(targets.map(node => this.reconcile(node, memories, options)));
+		this.sync(memories);
+	}
+
+	/**
+	 * Coalescing queue: if a doc is already reconciling, mark it dirty and abort the in-flight
+	 * request. The loop below always re-reads node.content fresh, so nothing is ever dropped.
+	 */
+	private reconcile(node: Memory, memories: Memory[] | MemoryCache, options: LLMRequest): Promise<void> {
+		const key = node.name;
+		const existing = this.queues.get(key);
+		if (existing) {
+			existing.dirty = true;
+			existing.request?.abort?.();
+			return existing.task;
+		}
+
+		const entry = {dirty: false, request: null, task: Promise.resolve()};
+		this.queues.set(key, entry);
+		const mem = this.unwrap(memories);
+		entry.task = (async () => {
+			do {
+				entry.dirty = false;
+				await this.reconcileDoc(node, mem, options, entry);
+			} while (entry.dirty);
+		})().finally(() => {
+			this.queues.delete(key);
+			rebuildGraph(mem);
+			this.sync(memories);
+		});
+		return entry.task;
+	}
+
+	private async reconcileDoc(node: Memory, memories: Memory[], options: LLMRequest, entry: {request: {abort?: () => void} | null}): Promise<void> {
+		const currentBody = this.stripHeader(node.content);
+		let update;
+		try {
+			for (let i = 0; i < 2 && !update?.content; i++) {
+				const request = this.llm.ask(currentBody, {
+					model: options.model,
+					temperature: 0.3,
+					schema: {
+						description: {type: 'string', description: 'One-line description of what this document covers, no formatting or emojis', required: true},
+						content: {type: 'string', description: 'Rewritten document body in markdown, without the frontmatter block', required: true},
+					},
+					system: `You are a knowledge base editor maintaining one document in an Obsidian-style vault.
+
+If the document has a "${FACTS_HEADING}" section, integrate every bullet under it into the appropriate part of the document, then remove the "${FACTS_HEADING}" section entirely. If there is no such section, just tidy the document per the rules below.
+
+Structure: follow this generic shape loosely, adapting section names/order to what the content actually needs (e.g. journal-style docs may want a timeline instead of "Details"):
+\`\`\`markdown
+${GENERIC_TEMPLATE}
+\`\`\`
+
+Formatting rules:
+- Use Obsidian-style markdown: # headings, **bold** for emphasis, bullet & numbered lists for grouped 1D data, tables for 2D data
+- Link related concepts with [[WikiLink]] notation using full paths like [[People/Sarah]] or [[Projects/Website]]
+- Create links for specific entities (person, place, project, program) and abstract concepts, but skip generics (car, red, dog)
+- Keep the document concise, factual, and human-readable
+- Resolve contradictions: newer facts always win — delete the outdated statement entirely, never keep both
+- Do not add frontmatter blocks, filler, preamble, or AI commentary
+
+Other nodes in the vault (link to these instead of duplicating their content):
+${this.listNodes(memories).filter(n => n.name !== node.name).map(n => n.name).join(', ') || 'none'}
+
+Current document:
+\`\`\`markdown
+${currentBody}
+\`\`\``,
+				});
+				entry.request = request;
+				update = await request;
+			}
+		} catch (err: any) {
+			if (err?.name === 'AbortError') return;
+			throw err;
+		} finally {
+			entry.request = null;
+		}
+
+		if (!update?.content) return;
+		node.description = node.name !== 'People/User' ? update.description : 'All information about the current user';
+		node.content = this.touchHeader(node, update.content);
+		const [e] = await this.llm.embedding(node.content);
+		if (e) node.embedding = e.embedding;
+	}
+
+	private async factAgent(conversation: string, memories: Memory[], options: LLMRequest, weekKey: string): Promise<FactBucket[]> {
+		const buckets = new Map<string, string[]>();
+		await this.llm.ask(conversation, {
+			model: options.model,
+			temperature: 0.2,
+			system: `You are a fact extractor. Analyze this conversation and extract facts worth remembering long-term.
+
+Rules:
+- ONLY extract current facts the USER explicitly stated about themselves, their work, or their projects
+- ONLY extract decisions that were MADE during this conversation
+- DO NOT extract anything the AI said, its capabilities, or meta-conversation about the AI
+- DO NOT extract greetings, pleasantries, or generic exchanges
+- DO NOT extract deltas or changes in facts; ONLY the end fact
+- If nothing worth remembering was said, do not call any tools
+
+When extracting facts, you MUST also decide the exact destination path:
+- Use an existing node name if the facts clearly belong there
+- All information primarily about the user should go under "People/User"
+- When required, create a new path following collection/subject format (e.g., People/Sarah, Projects/Oxide) — you are not limited to any fixed list of collections, use whatever fits
+- For journal entries, use "Journal"
+
+Available nodes:
+- Journal
+${this.listNodes(memories).filter(n => !n.name.includes('Journal')).map(n => `- ${n.name}: ${n.description}`).join('\n') || 'None yet.'}`,
+			tools: [{
+				name: 'facts_extract',
+				description: 'Submit facts with their destination',
+				args: {
+					destination: {type: 'string', description: 'Exact existing node name OR new path (e.g. "People/Sarah", "Projects/Oxide")', required: true},
+					facts: {type: 'string', description: 'Comma-separated facts', required: true},
+				},
+				fn: (args: any) => {
+					const subject = args.destination.trim().toLowerCase() === 'journal'
+						? `Journal/${weekKey}` : args.destination.trim();
+					const facts = buckets.get(subject) ?? [];
+					facts.push(...dedupeFacts(String(args.facts).split(',')));
+					buckets.set(subject, facts);
+					return 'Recorded';
+				},
+			}],
+		});
+		return buckets.entries().toArray().map(([subject, facts]) => ({subject, facts}));
+	}
+}

src/ollama.ts

@@ -1,113 +0,0 @@
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
-import {Ai} from './ai.ts';
-import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
-import {Ollama as ollama} from 'ollama';
-
-export class Ollama extends LLMProvider {
-	client!: ollama;
-
-	constructor(public readonly ai: Ai, public host: string, public model: string) {
-		super();
-		this.client = new ollama({host});
-	}
-
-	private toStandard(history: any[]): LLMMessage[] {
-		for(let i = 0; i < history.length; i++) {
-			if(history[i].role == 'assistant' && history[i].tool_calls) {
-				if(history[i].content) delete history[i].tool_calls;
-				else {
-					history.splice(i, 1);
-					i--;
-				}
-			} else if(history[i].role == 'tool') {
-				const error = history[i].content.startsWith('{"error":');
-				history[i] = {role: 'tool', name: history[i].tool_name, args: history[i].args, [error ? 'error' : 'content']: history[i].content};
-			}
-		}
-		return history;
-	}
-
-	private fromStandard(history: LLMMessage[]): any[] {
-		return history.map((h: any) => {
-			if(h.role != 'tool') return h;
-			return {role: 'tool', tool_name: h.name, content: h.error || h.content}
-		});
-	}
-
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
-		const controller = new AbortController();
-		const response = new Promise<any>(async (res, rej) => {
-			let system = options.system || this.ai.options.system;
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
-			if(history[0].roll == 'system') {
-				if(!system) system = history.shift();
-				else history.shift();
-			}
-			if(options.compress) history = await this.ai.llm.compress(<any>history, options.compress.max, options.compress.min);
-			if(options.system) history.unshift({role: 'system', content: system})
-
-			const requestParams: any = {
-				model: options.model || this.model,
-				messages: history,
-				stream: !!options.stream,
-				signal: controller.signal,
-				options: {
-					temperature: options.temperature || this.ai.options.temperature || 0.7,
-					num_predict: options.max_tokens || this.ai.options.max_tokens || 4096,
-				},
-				tools: (options.tools || this.ai.options.tools || []).map(t => ({
-					type: 'function',
-					function: {
-						name: t.name,
-						description: t.description,
-						parameters: {
-							type: 'object',
-							properties: t.args ? objectMap(t.args, (key, value) => ({...value, required: undefined})) : {},
-							required: t.args ? Object.entries(t.args).filter(t => t[1].required).map(t => t[0]) : []
-						}
-					}
-				}))
-			}
-
-			// Run tool chains
-			let resp: any;
-			do {
-				resp = await this.client.chat(requestParams);
-				if(options.stream) {
-					resp.message = {role: 'assistant', content: '', tool_calls: []};
-					for await (const chunk of resp) {
-						if(controller.signal.aborted) break;
-						if(chunk.message?.content) {
-							resp.message.content += chunk.message.content;
-							options.stream({text: chunk.message.content});
-						}
-						if(chunk.message?.tool_calls) resp.message.tool_calls = chunk.message.tool_calls;
-						if(chunk.done) break;
-					}
-				}
-
-				// Run tools
-				if(resp.message?.tool_calls?.length && !controller.signal.aborted) {
-					history.push(resp.message);
-					const results = await Promise.all(resp.message.tool_calls.map(async (toolCall: any) => {
-						const tool = (options.tools || this.ai.options.tools)?.find(findByProp('name', toolCall.function.name));
-						if(!tool) return {role: 'tool', tool_name: toolCall.function.name, content: '{"error": "Tool not found"}'};
-						const args = typeof toolCall.function.arguments === 'string' ? JSONAttemptParse(toolCall.function.arguments, {}) : toolCall.function.arguments;
-						try {
-							const result = await tool.fn(args, this.ai);
-							return {role: 'tool', tool_name: toolCall.function.name, args, content: JSONSanitize(result)};
-						} catch (err: any) {
-							return {role: 'tool', tool_name: toolCall.function.name, args, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'})};
-						}
-					}));
-					history.push(...results);
-					requestParams.messages = history;
-				}
-			} while (!controller.signal.aborted && resp.message?.tool_calls?.length);
-			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {role: 'assistant', content: resp.message?.content}]));
-		});
-		return Object.assign(response, {abort: () => controller.abort()});
-	}
-}

src/open-ai.ts

@@ -1,39 +1,57 @@
 import {OpenAI as openAI} from 'openai';
-import {findByProp, objectMap, JSONSanitize, JSONAttemptParse} from '@ztimson/utils';
+import {findByProp, objectMap, JSONSanitize, JSONAttemptParse, clean, makeArray} from '@ztimson/utils';
-import {Ai} from './ai.ts';
+import {AbortablePromise, Ai} from './ai.ts';
 import {LLMMessage, LLMRequest} from './llm.ts';
-import {AbortablePromise, LLMProvider} from './provider.ts';
+import {LLMProvider} from './provider.ts';
+import {TokenPool} from './token-pool.ts';
+import {convertSchema} from './tools.ts';
 
 export class OpenAi extends LLMProvider {
-	client!: openAI;
+	tokenPool!: TokenPool;
+	private clients = new Map<string, openAI>();
 
-	constructor(public readonly ai: Ai, public readonly apiToken: string, public model: string) {
+	constructor(public readonly ai: Ai, public readonly host: string | null, public readonly token: string | string[], public model: string) {
 		super();
-		this.client = new openAI({apiKey: apiToken});
+		const tokens = makeArray(token).filter(Boolean);
+		this.tokenPool = new TokenPool(...(tokens.length ? tokens : [host ? 'ignored' : '']));
+	}
+
+	private getClient(token: string): openAI {
+		let client = this.clients.get(token);
+		if(!client) {
+			client = new openAI(clean({baseURL: this.host, apiKey: token || undefined}));
+			this.clients.set(token, client);
+		}
+		return client;
 	}
 
 	private toStandard(history: any[]): LLMMessage[] {
 		for(let i = 0; i < history.length; i++) {
 			const h = history[i];
 			if(h.role === 'assistant' && h.tool_calls) {
-				const tools = h.tool_calls.map((tc: any) => ({
+				const items: any[] = [];
+				if(h.content) items.push({role: 'assistant', content: h.content, timestamp: h.timestamp, duration: h.duration, tps: h.tps});
+				items.push(...h.tool_calls.map((tc: any) => ({
 					role: 'tool',
 					id: tc.id,
 					name: tc.function.name,
-					args: JSONAttemptParse(tc.function.arguments, {})
+					args: JSONAttemptParse(tc.function.arguments, {}),
-				}));
+					timestamp: h.timestamp,
-				history.splice(i, 1, ...tools);
+					duration: h.duration,
-				i += tools.length - 1;
+					tps: h.tps
-			} else if(h.role === 'tool' && h.content) {
+				})));
+				history.splice(i, 1, ...items);
+				i += items.length - 1;
+			} else if(h.role === 'tool') {
 				const record = history.find(h2 => h.tool_call_id == h2.id);
 				if(record) {
-					if(h.content.includes('"error":')) record.error = h.content;
+					if(h.content?.includes('"error":')) record.error = h.content;
-					else record.content = h.content;
+					else record.content = h.content || '';
 				}
 				history.splice(i, 1);
 				i--;
 			}
-
+			if(!history[i]?.timestamp) history[i].timestamp = Date.now();
 		}
 		return history;
 	}
@@ -47,10 +65,12 @@ export class OpenAi extends LLMProvider {
 					tool_calls: [{ id: h.id, type: 'function', function: { name: h.name, arguments: JSON.stringify(h.args) } }],
 					refusal: null,
 					annotations: [],
+					timestamp: h.timestamp,
 				}, {
 					role: 'tool',
 					tool_call_id: h.id,
-					content: h.error || h.content
+					content: h.error || h.content,
+					timestamp: h.timestamp,
 				});
 			} else {
 				result.push(h);
@@ -59,19 +79,23 @@ export class OpenAi extends LLMProvider {
 		}, [] as any[]);
 	}
 
-	ask(message: string, options: LLMRequest = {}): AbortablePromise<LLMMessage[]> {
+	ask(message: string, options: LLMRequest = {}): AbortablePromise<string | any> {
 		const controller = new AbortController();
-		const response = new Promise<any>(async (res, rej) => {
+		return Object.assign(new Promise<any>(async (res, rej) => {
-			let history = this.fromStandard([...options.history || [], {role: 'user', content: message}]);
+			const base = (options.history || []).filter(h => h.role !== 'system');
-			if(options.compress) history = await this.ai.llm.compress(<any>history, options.compress.max, options.compress.min, options);
+			let history = this.fromStandard([
-
+				...(options.system ? [{role: <any>'system', content: options.system, timestamp: Date.now()}] : []),
+				...base,
+				{role: 'user', content: message, timestamp: Date.now()}
+			]);
+			const tools = options.tools || this.ai.options.llm?.tools || [];
 			const requestParams: any = {
 				model: options.model || this.model,
 				messages: history,
 				stream: !!options.stream,
-				max_tokens: options.max_tokens || this.ai.options.max_tokens || 4096,
+				max_completion_tokens: options.max_tokens || this.ai.options.llm?.max_tokens || undefined,
-				temperature: options.temperature || this.ai.options.temperature || 0.7,
+				temperature: options.temperature || this.ai.options.llm?.temperature || undefined,
-				tools: (options.tools || this.ai.options.tools || []).map(t => ({
+				tools: tools.map(t => ({
 					type: 'function',
 					function: {
 						name: t.name,
@@ -85,46 +109,110 @@ export class OpenAi extends LLMProvider {
 				}))
 			};
 
-			// Tool call and streaming logic similar to other providers
+			if(options.schema) {
-			let resp: any;
+				const schema = convertSchema(options.schema);
-			do {
+				requestParams.response_format = {
-				resp = await this.client.chat.completions.create(requestParams);
+					type: 'json_schema',
+					json_schema: {
+						name: 'response',
+						strict: true,
+						schema
+					}
+				};
+			}
 
-				// Implement streaming and tool call handling
+			if(options.stream) requestParams.stream_options = {include_usage: true};
+			let resp: any, terminal = false, duration = 0, tps = 0;
+			do {
+				requestParams.messages = history.map(({timestamp, ...m}) => m);
+				const callStart = Date.now();
+				resp = await this.tokenPool.run(token => this.getClient(token).chat.completions.create(requestParams)).catch(err => {
+					err.message += `\n\nMessages:\n${JSON.stringify(history, null, 2)}`;
+					throw err;
+				});
+
+				let usage: any;
 				if(options.stream) {
-					resp.choices = [];
+					resp.choices = [{message: {role: 'assistant', content: '', tool_calls: [], timestamp: Date.now()}}];
 					for await (const chunk of resp) {
 						if(controller.signal.aborted) break;
-						if(chunk.choices[0].delta.content) {
+						if(chunk.usage) usage = chunk.usage;
+						if(chunk.choices[0]?.delta?.content) {
+							resp.choices[0].message.content += chunk.choices[0].delta.content;
 							options.stream({text: chunk.choices[0].delta.content});
 						}
+						if(chunk.choices[0]?.delta?.tool_calls) {
+							for(const deltaTC of chunk.choices[0].delta.tool_calls) {
+								const existing = resp.choices[0].message.tool_calls.find(tc => tc.index === deltaTC.index);
+								if(existing) {
+									if(deltaTC.id) existing.id = deltaTC.id;
+									if(deltaTC.type) existing.type = deltaTC.type;
+									if(deltaTC.function) {
+										if(!existing.function) existing.function = {};
+										if(deltaTC.function.name) existing.function.name = deltaTC.function.name;
+										if(deltaTC.function.arguments) existing.function.arguments = (existing.function.arguments || '') + deltaTC.function.arguments;
+									}
+								} else {
+									resp.choices[0].message.tool_calls.push({
+										index: deltaTC.index,
+										id: deltaTC.id || '',
+										type: deltaTC.type || 'function',
+										function: {
+											name: deltaTC.function?.name || '',
+											arguments: deltaTC.function?.arguments || ''
+										}
+									});
+								}
+							}
+						}
 					}
+				} else {
+					usage = resp.usage;
 				}
+				duration = Date.now() - callStart;
+				tps = usage?.completion_tokens && duration > 0 ? usage.completion_tokens / (duration / 1000) : 0;
 
-				// Run tools
+				if(resp.error) throw new Error(resp.error);
 				const toolCalls = resp.choices[0].message.tool_calls || [];
 				if(toolCalls.length && !controller.signal.aborted) {
-					history.push(resp.choices[0].message);
+					history.push({...resp.choices[0].message, duration, tps});
 					const results = await Promise.all(toolCalls.map(async (toolCall: any) => {
-						const tool = options.tools?.find(findByProp('name', toolCall.function.name));
+						const tool = tools?.find(findByProp('name', toolCall.function.name));
-						if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}'};
+						if(options.stream) options.stream({tool: toolCall.function.name});
+						if(!tool) return {role: 'tool', tool_call_id: toolCall.id, content: '{"error": "Tool not found"}', timestamp: Date.now()};
 						try {
 							const args = JSONAttemptParse(toolCall.function.arguments, {});
-							const result = await tool.fn(args, this.ai);
+							const toolStream = options.stream && ((chunk: any) => {
-							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize(result)};
+								if(chunk.done) { terminal = true; return; }
+								options.stream!(chunk);
+							});
+							const result = await tool.fn(args, toolStream, this.ai, toolCall.id);
+							return {role: 'tool', tool_call_id: toolCall.id, content: typeof result == 'object' ? JSONSanitize(result) : result, timestamp: Date.now()};
 						} catch (err: any) {
-							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'})};
+							return {role: 'tool', tool_call_id: toolCall.id, content: JSONSanitize({error: err?.message || err?.toString() || 'Unknown'}), timestamp: Date.now()};
 						}
 					}));
 					history.push(...results);
 					requestParams.messages = history;
 				}
-			} while (!controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
+			} while (!terminal && !controller.signal.aborted && resp.choices?.[0]?.message?.tool_calls?.length);
 
+			if(!terminal) {
+				const textContent = resp.choices[0].message.content || '';
+				history.push({role: 'assistant', content: textContent.trim(), timestamp: Date.now(), duration, tps});
+			}
+
+			history = this.toStandard(history);
+			if(options.history) options.history.splice(0, options.history.length, ...history.filter(h => h.role !== 'system'));
 			if(options.stream) options.stream({done: true});
-			res(this.toStandard([...history, {role: 'assistant', content: resp.choices[0].message.content || ''}]));
-		});
 
-		return Object.assign(response, {abort: () => controller.abort()});
+			const turnStart = history.map(h => h.role).lastIndexOf('user');
+			const finalContent = history.slice(turnStart + 1).reduce((str, h) => {
+				if(h.role === 'assistant') return str + (h.content || '');
+				return str;
+			}, '').trim();
+
+			res(options.schema ? JSONAttemptParse(finalContent, finalContent) : finalContent);
+		}), {abort: () => controller.abort()});
 	}
 }

src/provider.ts

@@ -1,7 +1,6 @@
-import {LLMMessage, LLMOptions, LLMRequest} from './llm.ts';
+import {AbortablePromise} from './ai.ts';
-
+import {LLMRequest} from './llm.ts';
-export type AbortablePromise<T> = Promise<T> & {abort: () => void};
 
 export abstract class LLMProvider {
-	abstract ask(message: string, options: LLMRequest): AbortablePromise<LLMMessage[]>;
+	abstract ask(message: string, options: LLMRequest): AbortablePromise<string>;
 }

src/token-pool.ts

@@ -0,0 +1,65 @@
+const DEFAULT_COOLDOWN = 15 * 60 * 1000;
+
+type TokenState = {
+	token: string;
+	cooldownUntil: number; // 0 = available now
+	lastError?: {code: number, message: string};
+};
+
+export class TokenPoolExhaustedError extends Error {
+	constructor(public tokens: Record<string, {code: number, message: string}>) {
+		super(`All tokens exhausted:\n${Object.entries(tokens).map(([t, e]) => `${t}: [${e.code}] ${e.message}`).join('\n')}`);
+		this.name = 'TokenPoolExhaustedError';
+	}
+}
+
+export class TokenPool {
+	private states: TokenState[];
+
+	constructor(...tokens: string[]) {
+		this.states = tokens.map(token => ({token, cooldownUntil: 0}));
+	}
+
+	private preview(token: string): string {
+		return token.length <= 8 ? '****' : `${token.slice(0, 4)}...${token.slice(-4)}`;
+	}
+
+	/** Anthropic & OpenAI SDKs both attach `status` to thrown errors */
+	private statusCode(err: any): number {
+		return err?.status ?? err?.response?.status ?? err?.statusCode;
+	}
+
+	private retryAfter(err: any): number {
+		const headers = err?.headers || err?.response?.headers;
+		const raw = headers?.get?.('retry-after') ?? headers?.['retry-after'];
+		if(raw) {
+			const seconds = Number(raw);
+			if(!isNaN(seconds)) return Date.now() + seconds * 1000;
+			const date = new Date(raw).getTime();
+			if(!isNaN(date)) return date;
+		}
+		return Date.now() + DEFAULT_COOLDOWN;
+	}
+
+	async run<T>(fn: (token: string) => Promise<T>): Promise<T> {
+		const now = Date.now();
+		for(const state of this.states) {
+			if(state.cooldownUntil > now) continue;
+			try {
+				const result = await fn(state.token);
+				state.cooldownUntil = 0;
+				state.lastError = undefined;
+				return result;
+			} catch(err: any) {
+				const code = this.statusCode(err);
+				if(![401, 403, 429].includes(code)) throw err;
+				state.cooldownUntil = code === 429 ? this.retryAfter(err) : Date.now() + DEFAULT_COOLDOWN;
+				state.lastError = {code, message: err?.message || 'Unknown error'};
+			}
+		}
+
+		const failures: Record<string, {code: number, message: string}> = {};
+		this.states.forEach(s => { if(s.lastError) failures[this.preview(s.token)] = s.lastError; });
+		throw new TokenPoolExhaustedError(failures);
+	}
+}

src/tools.ts

@@ -1,6 +1,16 @@
-import {$, $Sync} from '@ztimson/node-utils';
+import  * as cheerio from 'cheerio';
-import {ASet, consoleInterceptor, Http, fn as Fn} from '@ztimson/utils';
+import {$Sync} from '@ztimson/node-utils';
+import {ASet, consoleInterceptor, Http, fn as Fn, decodeHtml, objectMap} from '@ztimson/utils';
+import * as os from 'node:os';
 import {Ai} from './ai.ts';
+import {LLMRequest} from './llm.ts';
+
+const UA = 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)';
+
+const getShell = () => {
+	if(os.platform() == 'win32') return 'cmd';
+	return $Sync`echo $SHELL`?.split('/').pop() || 'bash';
+}
 
 export type AiToolArg = {[key: string]: {
 	/** Argument type */
@@ -31,40 +41,103 @@ export type AiTool = {
 	/** Tool arguments */
 	args?: AiToolArg,
 	/** Callback function */
-	fn: (args: any, ai: Ai) => any | Promise<any>,
+	fn: (args: any, stream: LLMRequest['stream'], ai: Ai, toolId?: string) => any | Promise<any>,
 };
 
-export const CliTool: AiTool = {
+export function convertSchema(schema: any): any {
+	if(!schema) return null;
+
+	const convertProp = (prop: any): any => {
+		const converted: any = {
+			type: prop.type || 'string',
+		};
+
+		if(prop.description) converted.description = prop.description;
+		if(prop.default !== undefined) converted.default = prop.default;
+		if(prop.enum) converted.enum = prop.enum;
+		if(prop.pattern) converted.pattern = prop.pattern;
+
+		// Handle array items
+		if(prop.type === 'array' && prop.items) {
+			converted.items = convertProp(prop.items);
+		}
+
+		// Handle object properties
+		if(prop.type === 'object' && prop.items) {
+			converted.properties = objectMap(prop.items, (key, value) => convertProp(value));
+			const required = Object.entries(prop.items).filter(([_, v]: any) => v.required).map(([k]) => k);
+			if(required.length) converted.required = required;
+			converted.additionalProperties = false;
+		}
+
+		// Handle min/max based on type
+		if(prop.min !== undefined) {
+			if(prop.type === 'string' || prop.type === 'array') converted.minLength = prop.min;
+			else converted.minimum = prop.min;
+		}
+		if(prop.max !== undefined) {
+			if(prop.type === 'string' || prop.type === 'array') converted.maxLength = prop.max;
+			else converted.maximum = prop.max;
+		}
+
+		return converted;
+	};
+
+	return {
+		type: 'object',
+		properties: objectMap(schema, (key, value) => convertProp(value)),
+		required: Object.entries(schema).filter(([_, v]: any) => v.required).map(([k]) => k),
+		additionalProperties: false
+	};
+}
+
+export const ExecCliTool: AiTool = {
 	name: 'cli',
 	description: 'Use the command line interface, returns any output',
 	args: {command: {type: 'string', description: 'Command to run', required: true}},
-	fn: (args: {command: string}) => $`${args.command}`
+	fn: (args: {command: string}) => $Sync`${args.command}`
 }
 
-export const DateTimeTool: AiTool = {
+export const ExecJSTool: AiTool = {
-	name: 'get_datetime',
+	name: 'exec_javascript',
-	description: 'Get current date and time',
+	description: 'Execute commonjs javascript',
-	args: {},
+	args: {
-	fn: async () => new Date().toISOString()
+		code: {type: 'string', description: 'CommonJS javascript', required: true}
+	},
+	fn: async (args: {code: string}) => {
+		const c = consoleInterceptor(null);
+		const resp = await Fn<any>({console: c}, args.code, true).catch((err: any) => c.output.error.push(err));
+		return {...c.output, return: resp, stdout: undefined, stderr: undefined};
+	}
+}
+
+export const ExecPythonTool: AiTool = {
+	name: 'exec_python',
+	description: 'Execute commonjs javascript',
+	args: {
+		code: {type: 'string', description: 'CommonJS javascript', required: true}
+	},
+	fn: async (args: {code: string}) => ({result: $Sync`python -c "${args.code}"`})
 }
 
 export const ExecTool: AiTool = {
 	name: 'exec',
 	description: 'Run code/scripts',
 	args: {
-		language: {type: 'string', description: 'Execution language', enum: ['cli', 'node', 'python'], required: true},
+		language: {type: 'string', description: `Execution language (CLI: ${getShell()})`, enum: ['cli', 'node', 'python'], required: true},
 		code: {type: 'string', description: 'Code to execute', required: true}
 	},
-	fn: async (args, ai) => {
+	fn: async (args, stream, ai) => {
 		try {
-			switch(args.type) {
+			switch(args.language) {
-				case 'bash':
+				case 'cli':
-					return await CliTool.fn({command: args.code}, ai);
+					return await ExecCliTool.fn({command: args.code}, stream, ai);
 				case 'node':
-					return await JSTool.fn({code: args.code}, ai);
+					return await ExecJSTool.fn({code: args.code}, stream, ai);
-				case 'python': {
+				case 'python':
-					return await PythonTool.fn({code: args.code}, ai);
+					return await ExecPythonTool.fn({code: args.code}, stream, ai);
-				}
+				default:
+					throw new Error(`Unsupported language: ${args.language}`);
 			}
 		} catch(err: any) {
 			return {error: err?.message || err.toString()};
@@ -72,8 +145,483 @@ export const ExecTool: AiTool = {
 	}
 }
 
-export const FetchTool: AiTool = {
+export const FsDeleteTool = (whitelist: null | string[] = null): AiTool => {
-	name: 'fetch',
+	return {
+		name: 'fs_delete',
+		description: 'Delete a file or directory',
+		args: {
+			path: {type: 'string', description: 'Path to file or directory', required: true},
+			recursive: {type: 'boolean', description: 'Delete all children', required: false}
+		},
+		fn: async ({path, recursive = false}) => {
+			const {existsSync, rmSync} = await import('fs');
+			const normalizePath = p => p.replace(/\\/g, '/');
+
+			path = normalizePath(path);
+			if(whitelist && !whitelist.some(p => path.startsWith(p))) return {error: 'Permission denied'};
+			if(!existsSync(path)) return {error: 'Path does not exist'};
+
+			rmSync(path, {recursive, force: true});
+			return {success: true, path};
+		}
+	}
+}
+
+export const FsMoveTool = (whitelist: null | string[] = null): AiTool => {
+	return {
+		name: 'fs_move',
+		description: 'Move or rename a file or directory',
+		args: {
+			source: {type: 'string', description: 'Path to source file or directory', required: true},
+			destination: {type: 'string', description: 'Path to destination file or directory', required: true}
+		},
+		fn: async ({source, destination}) => {
+			const {existsSync, renameSync} = await import('fs');
+			const normalizePath = p => p.replace(/\\/g, '/');
+
+			source = normalizePath(source);
+			destination = normalizePath(destination);
+			if(whitelist && !whitelist.some(p => source.startsWith(p) && destination.startsWith(p))) return {error: 'Permission denied'};
+
+			if(!existsSync(source)) return {error: 'Source path does not exist'};
+			if(existsSync(destination)) return {error: 'Destination path already exists'};
+
+			renameSync(source, destination);
+			return {success: true, source, destination};
+		}
+	}
+}
+
+export const FsReadTool = (whitelist: null | string[] = null): AiTool => {
+	return {
+		name: 'fs_read',
+		description: 'Read the contents of a provided path. Works with files and directories',
+		args: {path: {type: 'string', description: 'Path to file or directory', required: true}},
+		fn: async ({path}) => {
+			const {existsSync, lstatSync, readdirSync, readFileSync} = await import('fs');
+			const {join} = await import('path');
+			const normalizePath = p => p.replace(/\\/g, '/');
+
+			path = normalizePath(path);
+			if(whitelist && !whitelist.some(p => path.startsWith(p))) return {error: 'Permission denied'};
+
+			if(!existsSync(path)) return {error: 'Path does not exist'};
+			const stats = lstatSync(path);
+			if(stats.isDirectory()) {
+				const children = readdirSync(path).map(name => {
+					const childPath = normalizePath(join(path, name));
+					const childStats = lstatSync(childPath);
+					return {name, type: childStats.isDirectory() ? 'directory' : 'file', size: childStats.size};
+				});
+				return {type: 'directory', children};
+			}
+			const content = readFileSync(path, 'utf-8');
+			return {type: 'file', content};
+		}
+	}
+}
+
+export const FsSearchTool = (whitelist: null | string[] = null): AiTool => {
+	return {
+		name: 'fs_search',
+		description: 'Scan a directory for matching glob patterns (e.g. "**/*.js", "src/**/*.test.ts")',
+		args: {
+			pattern: {type: 'string', description: 'Glob pattern to match against paths', required: true},
+			root: {type: 'string', description: 'Directory to search from', required: false, default: '.'}
+		},
+		fn: async ({pattern, root = '.'}) => {
+			const {existsSync, lstatSync, readdirSync} = await import('fs');
+			const {join, relative} = await import('path');
+			const normalizePath = p => p.replace(/\\/g, '/');
+
+			root = normalizePath(root);
+			if(!existsSync(root)) return {error: 'Root path does not exist'};
+			if(!lstatSync(root).isDirectory()) return {error: 'Root path is not a directory'};
+
+			if(whitelist && !whitelist.some(p => root.startsWith(p))) return {error: 'Permission denied'};
+
+			const globToRegex = (glob) => {
+				let re = '';
+				for(let i = 0; i < glob.length; i++) {
+					const c = glob[i];
+					if(c === '*') {
+						if(glob[i + 1] === '*') {
+							const isSlash = glob[i + 2] === '/';
+							re += '.*';
+							i += isSlash ? 2 : 1;
+						} else {
+							re += '[^/]*';
+						}
+					} else if(c === '?') {
+						re += '[^/]';
+					} else if('.+^$(){}|[]\\'.includes(c)) {
+						re += '\\' + c;
+					} else {
+						re += c;
+					}
+				}
+				return new RegExp('^' + re + '$');
+			};
+			const regex = globToRegex(pattern);
+
+			const results: any = [];
+			const walk = (dir) => {
+				for(const name of readdirSync(dir)) {
+					const fullPath = normalizePath(join(dir, name));
+					const stats = lstatSync(fullPath);
+					const relPath = normalizePath(relative(root, fullPath));
+					if(regex.test(relPath)) {
+						results.push({path: relPath, type: stats.isDirectory() ? 'directory' : 'file', size: stats.size});
+					}
+					if(stats.isDirectory()) walk(fullPath);
+				}
+			};
+			walk(root);
+
+			return results;
+		}
+	}
+}
+
+export const FsWriteTool = (whitelist: null | string[] = null): AiTool => {
+	return {
+		name: 'fs_write',
+		description: 'Create a directory, write content to a file or preform a find & replace',
+		args: {
+			path: {type: 'string', description: 'Path to file or directory', required: true},
+			content: {type: 'string', description: 'Content to write or replace (Omit to create a directory)'},
+			find: {type: 'string', description: 'Text or regex pattern to match (regex must match pattern: "/pattern/g")'}
+		},
+		fn: async ({path, content, find}) => {
+			const {existsSync, mkdirSync, readFileSync, writeFileSync} = await import('fs');
+			const {dirname} = await import('path');
+			const normalizePath = p => p.replace(/\\/g, '/');
+
+			path = normalizePath(path);
+			if(whitelist && !whitelist.some(p => path.startsWith(p))) return {error: 'Permission denied'};
+
+			if(content === undefined) {
+				mkdirSync(path, {recursive: true});
+				return {success: true, type: 'directory', path};
+			}
+
+			const dir = normalizePath(dirname(path));
+			if(!existsSync(dir)) mkdirSync(dir, {recursive: true});
+
+			if(find && existsSync(path)) {
+				const existing = readFileSync(path, 'utf-8');
+				const regexMatch = find.match(/^\/(.+)\/([gimuy]*)$/);
+				const pattern = regexMatch ? new RegExp(regexMatch[1], regexMatch[2]) : find;
+
+				if(!existing.match(pattern)) return {error: 'Find pattern not found in file'};
+
+				const updated = existing.replace(pattern, content);
+				writeFileSync(path, updated, 'utf-8');
+				return {success: true, type: 'file', path, replaced: true, content: updated};
+			}
+
+			writeFileSync(path, content, 'utf-8');
+			return {success: true, type: 'file', path, content};
+		}
+	}
+}
+
+export const GetPathsTool: AiTool = {
+	name: 'get_paths',
+	description: 'Get the current working directory, and paths to the users home directory',
+	fn: async () => {
+		return {
+			home: os.homedir(),
+			cwd: process.cwd()
+		};
+	}
+}
+
+export const GetDatetimeTool: AiTool = {
+	name: 'get_datetime',
+	description: 'Get local/UTC timestamp',
+	args: {
+		timezone: {type: 'string', description: 'Which timezone to return, defaults to local', enum: ['local', 'utc'], default: 'local'}
+	},
+	fn: ({timezone}) => new Date()[timezone === 'local' ? 'toString' : 'toUTCString']()
+}
+
+export const GetDevice: AiTool = {
+	name: 'get_device',
+	description: 'Get comprehensive system information including hostname, specs, load, storage, and network status',
+	args: {},
+	fn: async () => {
+		const platform = os.platform();
+		const hostname = os.hostname();
+
+		// CPU Info
+		const cpus = os.cpus();
+		const cpuModel = cpus[0].model;
+		const cpuCores = cpus.length;
+
+		// Memory Info
+		const totalMem: any = (os.totalmem() / 1024 / 1024 / 1024).toFixed(2);
+		const freeMem: any = (os.freemem() / 1024 / 1024 / 1024).toFixed(2);
+		const usedMem: any = (totalMem - freeMem).toFixed(2);
+		const memUsage: any = ((usedMem / totalMem) * 100).toFixed(1);
+
+		// Load Average (not available on Windows)
+		const loadAvg = platform === 'win32' ? ['N/A', 'N/A', 'N/A'] : os.loadavg().map(l => l.toFixed(2));
+
+		// Storage Usage
+		let storage = {};
+		if(platform === 'win32') {
+			const ps = $Sync`powershell "Get-PSDrive C | Select-Object Used,Free | ConvertTo-Json"`.trim();
+			const drive = JSON.parse(ps);
+			const used: any = (drive.Used / 1024 / 1024 / 1024).toFixed(2);
+			const free: any = (drive.Free / 1024 / 1024 / 1024).toFixed(2);
+			const total: any = (parseFloat(used) + parseFloat(free)).toFixed(2);
+			const usage: any = ((used / total) * 100).toFixed(1);
+			storage = {
+				filesystem: 'C:',
+				size: `${total} GB`,
+				used: `${used} GB`,
+				available: `${free} GB`,
+				usage: `${usage}%`
+			};
+		} else {
+			const df = $Sync`df -h / | tail -1`.trim();
+			const s = df.split(/\s+/);
+			storage = {
+				filesystem: s[0],
+				size: s[1],
+				used: s[2],
+				available: s[3],
+				usage: s[4]
+			};
+		}
+
+		// Network Status
+		const interfaces = os.networkInterfaces();
+		const activeIfaces = Object.entries(interfaces)
+			.filter(([name]) => name !== 'lo' && !name.includes('Loopback'))
+			.map(([name, addrs]) => {
+				const ipv4 = addrs?.find(a => a.family === 'IPv4');
+				return ipv4 ? {name, ip: ipv4.address} : null;
+			})
+			.filter(Boolean);
+
+		// Internet connectivity check
+		let internet = false;
+		try {
+			if(platform === 'win32') {
+				$Sync`powershell "Test-Connection -ComputerName 8.8.8.8 -Count 1 -Quiet"`;
+			} else {
+				$Sync`ping -c 1 -W 2 8.8.8.8 > /dev/null 2>&1`;
+			}
+			internet = true;
+		} catch {}
+
+		// Uptime
+		const uptime = os.uptime();
+		const days = Math.floor(uptime / 86400);
+		const hours = Math.floor((uptime % 86400) / 3600);
+		const minutes = Math.floor((uptime % 3600) / 60);
+
+		return {
+			hostname,
+			cpu: {
+				model: cpuModel,
+				cores: cpuCores
+			},
+			memory: {
+				total: `${totalMem} GB`,
+				used: `${usedMem} GB`,
+				free: `${freeMem} GB`,
+				usage: `${memUsage}%`
+			},
+			load: {
+				'1min': loadAvg[0],
+				'5min': loadAvg[1],
+				'15min': loadAvg[2]
+			},
+			storage,
+			network: {
+				interfaces: activeIfaces,
+				internet: internet ? 'connected' : 'disconnected'
+			},
+			uptime: `${days}d ${hours}h ${minutes}m`,
+			platform: `${os.type()} ${os.release()}`
+		};
+	}
+}
+
+export const GetWikipediaTool: AiTool = {
+	name: 'get_wikipedia',
+	description: 'Search Wikipedia for matching articles',
+	args: {
+		query: {type: 'string', description: 'Search term or article title', required: true},
+		mode: {type: 'string', description: 'search - look for articles, summary - intro of first found article (default), full - complete first found article', enum: ['search', 'summary', 'full'], default: 'summary'},
+		ua: {type: 'string', description: 'User Agent'},
+	},
+	fn: async ({query, mode, ua}) => {
+		class WikipediaClient {
+			useragent = 'Mozilla/5.0 (Windows NT 10.0; Win64; x64)';
+
+			constructor(useragent: string) {
+				this.useragent = useragent;
+			}
+
+			async get(url) {
+				const resp = await fetch(url, {headers: {'User-Agent': this.useragent}});
+				return resp.json();
+			}
+
+			api(params) {
+				const qs = new URLSearchParams({...params, format: 'json', utf8: '1'}).toString();
+				return this.get(`https://en.wikipedia.org/w/api.php?${qs}`);
+			}
+
+			clean(text) {
+				const cutoffs = ['== See also ==', '== References ==', '== Bibliography ==', '== External links =='];
+				for (const marker of cutoffs) {
+					const idx = text.indexOf(marker);
+					if (idx !== -1) text = text.slice(0, idx);
+				}
+
+				return text
+					.replace(/^={4}\s*(.+?)\s*={4}$/gm, '#### $1')
+					.replace(/^={3}\s*(.+?)\s*={3}$/gm, '### $1')
+					.replace(/^={2}\s*(.+?)\s*={2}$/gm, '## $1')
+					.replace(/\n{3,}/g, '\n\n')
+					.replace(/ {2,}/g, ' ')
+					.replace(/\[\d+]/g, '')
+					.trim();
+			}
+
+			async searchTitles(query: string, limit = 6) {
+				const data = await this.api({action: 'query', list: 'search', srsearch: query, srlimit: limit, srprop: 'snippet'});
+				return data.query?.search || [];
+			}
+
+			async fetchExtract(title: string, introOnly = false) {
+				const params: any = {action: 'query', prop: 'extracts', titles: title, explaintext: 1, redirects: 1};
+				if(introOnly) params.exintro = 1;
+				const data = await this.api(params);
+				const page: any = Object.values(data.query?.pages || {})[0];
+				return this.clean(page?.extract || '');
+			}
+
+			pageUrl(title: string) {
+				return `https://en.wikipedia.org/wiki/${encodeURIComponent(title.replace(/ /g, '_'))}`;
+			}
+
+			stripHtml(text: string) {
+				return text.replace(/<[^>]+>/g, '');
+			}
+
+			async lookup(query: string, detail = 'summary') {
+				const results = await this.searchTitles(query, 6);
+				if(!results.length) return `❌ No Wikipedia articles found for "${query}"`;
+				const title = results[0].title;
+				const url = this.pageUrl(title);
+				const introOnly = detail !== 'full';
+				const content = await this.fetchExtract(title, introOnly);
+				return `## ${title}\n🔗 ${url}\n\n${content}`;
+			}
+
+			async search(query: string) {
+				const results = await this.searchTitles(query, 8);
+				if(!results.length) return `❌ No results for "${query}"`;
+				const lines = [`### Search results for "${query}"\n`];
+				for(let i = 0; i < results.length; i++) {
+					const r = results[i];
+					const snippet = this.stripHtml(r.snippet || '').trim();
+					lines.push(`**${i + 1}. ${r.title}**\n${snippet}\n${this.pageUrl(r.title)}`);
+				}
+				return lines.join('\n\n');
+			}
+		}
+
+		const wiki = new WikipediaClient(ua);
+		if(mode === 'search') return wiki.search(query);
+		return wiki.lookup(query, mode || 'summary');
+	}
+};
+
+export const GeoCodeTool: AiTool = {
+	name: 'geo_code',
+	description: 'Converts coordinates to address OR vice versa',
+	args: {
+		query: {type: 'string', description: 'Search query - coordinates (lat,lon) or address string', required: true},
+	},
+	fn: async ({query}) => {
+		const coordinates = /(-?\d+(?:\.\d+)?).*?,.*?(-?\d+(?:\.\d+)?)/.exec(query);
+		if(coordinates) { // Geolocate
+			const url = `https://nominatim.openstreetmap.org/reverse?format=json&lat=${encodeURIComponent(coordinates[1])}&lon=${encodeURIComponent(coordinates[2])}`;
+			const response = await fetch(url, {headers: {'User-Agent': 'OpenSight/1.0', 'Accept-Language': 'en'}});
+			const data = await response.json();
+			if(data.display_name) return {address: data.display_name, mode: 'geolocate'};
+		} else { // Geocode
+			const url = `https://nominatim.openstreetmap.org/search?format=json&q=${encodeURIComponent(query)}`;
+			const response = await fetch(url, {headers: {'User-Agent': 'OpenSight/1.0'}});
+			const data = await response.json();
+			if(data[0]) return {latitude: parseFloat(data[0].lat), longitude: parseFloat(data[0].lon), mode: 'geocode'};
+		}
+		return {error: 'Not found'};
+	},
+}
+
+export const GeoWeatherTool: AiTool = {
+	name: 'geo_weather',
+	description: 'Gets weather and air quality info for a location and time',
+	args: {
+		query: {type: 'string', description: 'Location - address or place name', required: true},
+		day: {type: 'string', description: 'Date to retrieve (YYYY-MM-DD), defaults to today'},
+	},
+	fn: async ({query, day}) => {
+		day = day || new Date().toISOString().slice(0, 10);
+
+		const geoUrl = `https://nominatim.openstreetmap.org/search?format=json&q=${encodeURIComponent(query)}`;
+		const geoResponse = await fetch(geoUrl, {headers: {'User-Agent': 'OpenSight/1.0'}});
+		const geoData = await geoResponse.json();
+		if(!geoData[0]) return {error: 'Location not found'};
+
+		const lat = parseFloat(geoData[0].lat);
+		const lon = parseFloat(geoData[0].lon);
+
+		const weatherUrl = `https://api.open-meteo.com/v1/forecast?latitude=${lat}&longitude=${lon}&start_date=${day}&end_date=${day}&daily=weathercode,temperature_2m_max,temperature_2m_min,apparent_temperature_max,apparent_temperature_min,precipitation_sum,precipitation_probability_max,windspeed_10m_max,winddirection_10m_dominant,uv_index_max,sunrise,sunset&timezone=auto`;
+		const airUrl = `https://air-quality-api.open-meteo.com/v1/air-quality?latitude=${lat}&longitude=${lon}&start_date=${day}&end_date=${day}&hourly=us_aqi,european_aqi,pm10,pm2_5&timezone=auto`;
+
+		const [weatherResponse, airResponse] = await Promise.all([fetch(weatherUrl), fetch(airUrl)]);
+		const weatherData = await weatherResponse.json();
+		const airData = await airResponse.json();
+
+		const avg = arr => (arr && arr.length) ? arr.reduce((a, b) => a + b, 0) / arr.length : null;
+
+		return {
+			location: geoData[0].display_name,
+			latitude: lat,
+			longitude: lon,
+			elevation: weatherData.elevation,
+			date: day,
+			weatherCode: weatherData.daily?.weathercode?.[0],
+			tempMax: weatherData.daily?.temperature_2m_max?.[0],
+			tempMin: weatherData.daily?.temperature_2m_min?.[0],
+			feelsLikeMax: weatherData.daily?.apparent_temperature_max?.[0],
+			feelsLikeMin: weatherData.daily?.apparent_temperature_min?.[0],
+			precipitation: weatherData.daily?.precipitation_sum?.[0],
+			precipitationChance: weatherData.daily?.precipitation_probability_max?.[0],
+			windSpeedMax: weatherData.daily?.windspeed_10m_max?.[0],
+			windDirection: weatherData.daily?.winddirection_10m_dominant?.[0],
+			uvIndexMax: weatherData.daily?.uv_index_max?.[0],
+			sunrise: weatherData.daily?.sunrise?.[0],
+			sunset: weatherData.daily?.sunset?.[0],
+			usAqi: avg(airData.hourly?.us_aqi),
+			europeanAqi: avg(airData.hourly?.european_aqi),
+			pm10: avg(airData.hourly?.pm10),
+			pm2_5: avg(airData.hourly?.pm2_5),
+		};
+	},
+}
+
+export const WebFetchTool: AiTool = {
+	name: 'web_fetch',
 	description: 'Make HTTP request to URL',
 	args: {
 		url: {type: 'string', description: 'URL to fetch', required: true},
@@ -89,31 +637,164 @@ export const FetchTool: AiTool = {
 	}) => new Http({url: args.url, headers: args.headers}).request({method: args.method || 'GET', body: args.body})
 }
 
-export const JSTool: AiTool = {
+export const WebFlareSolverTool = (host: string) => {
-	name: 'exec_javascript',
+	return {
-	description: 'Execute commonjs javascript',
+		name: 'web_flaresolverr',
-	args: {
+		description: 'Use a flaresolverr proxy to bypass cloudflare bot detection',
-		code: {type: 'string', description: 'CommonJS javascript', required: true}
+		args: {
-	},
+			url: {type: 'string', description: 'URL to fetch', required: true},
-	fn: async (args: {code: string}) => {
+			cmd: {type: 'string', description: 'Flaresolverr cmd', enum: ['request.get', 'request.post'], default: 'request.get'},
-		const console = consoleInterceptor(null);
+			maxTimeout: {type: 'number', description: 'Fetch time limit', default: 60_000},
-		const resp = await Fn<any>({console}, args.code, true).catch((err: any) => console.output.error.push(err));
+			postData: {type: 'object', description: 'Data to send during request.post requests'},
-		return {...console.output, return: resp, stdout: undefined, stderr: undefined};
+		},
+		fn: async ({url, cmd, maxTimeout, postData}) => {
+			function toFormUrlEncoded(obj, prefix = '') {
+				const pairs: any = [];
+				for (const key in obj) {
+					if (!obj.hasOwnProperty(key)) continue;
+
+					const value = obj[key];
+					const encodedKey = prefix
+						? `${prefix}[${encodeURIComponent(key)}]`
+						: encodeURIComponent(key);
+
+					if (value === null || value === undefined) {
+						pairs.push(`${encodedKey}=`);
+					} else if (typeof value === 'object' && !Array.isArray(value)) {
+						pairs.push(toFormUrlEncoded(value, encodedKey));
+					} else if (Array.isArray(value)) {
+						value.forEach(item => {
+							pairs.push(`${encodedKey}[]=${encodeURIComponent(item)}`);
+						});
+					} else {
+						pairs.push(`${encodedKey}=${encodeURIComponent(value)}`);
+					}
+				}
+
+				return pairs.join('&');
+			}
+
+			const res = await fetch(host + '/v1', {
+				method: 'POST',
+				headers: {'Content-Type': 'application/json'},
+				body: JSON.stringify({cmd, url, maxTimeout, postData: postData ? toFormUrlEncoded(postData) : undefined}),
+			});
+
+			if(!res.ok) throw new Error(`FlareSolverr HTTP error: ${res.status} ${res.statusText}`);
+			const data = await res.json();
+			if(data.status !== 'ok') throw new Error(`FlareSolverr error: ${data.message ?? data.status}`);
+			return data.solution.response;
+		}
 	}
 }
 
-export const PythonTool: AiTool = {
+export const WebReadTool: AiTool = {
-	name: 'exec_javascript',
+	name: 'web_read',
-	description: 'Execute commonjs javascript',
+	description: 'Extract clean content from webpages, or convert media/documents to accessible formats',
 	args: {
-		code: {type: 'string', description: 'CommonJS javascript', required: true}
+		url: {type: 'string', description: 'URL to read', required: true},
+		mimeRegex: {type: 'string', description: 'Optional regex to filter MIME types (e.g., "^image/", "text/")'}
 	},
-	fn: async (args: {code: string}) => ({result: $Sync`python -c "${args.code}"`})
+	fn: async (args: {url: string; mimeRegex?: string}) => {
-}
+		const ua = 'AiTools-Webpage/1.0';
+		const maxSize = 10 * 1024 * 1024;
 
-export const SearchTool: AiTool = {
+		const response = await fetch(args.url, {
-	name: 'search',
+			headers: {
-	description: 'Use a search engine to find relevant URLs, should be changed with fetch to scrape sources',
+				'User-Agent': ua,
+				'Accept': 'text/html,application/xhtml+xml,application/xml;q=0.9,image/webp,*/*;q=0.8',
+				'Accept-Language': 'en-US,en;q=0.5'
+			},
+			redirect: 'follow'
+		}).catch(err => {throw new Error(`Failed to fetch: ${err.message}`)});
+
+		const contentType = response.headers.get('content-type') || '';
+		const mimeType = contentType.split(';')[0].trim().toLowerCase();
+
+		if(args.mimeRegex && !new RegExp(args.mimeRegex, 'i').test(mimeType)) {
+			return `❌ MIME type rejected: ${mimeType} (filter: ${args.mimeRegex})`;
+		}
+
+		if(mimeType.match(/^(image|audio|video)\//)) {
+			const buffer = await response.arrayBuffer();
+			if(buffer.byteLength > maxSize) {
+				return `❌ File too large: ${(buffer.byteLength / 1024 / 1024).toFixed(1)}MB (max 10MB)\nType: ${mimeType}`;
+			}
+			const base64 = Buffer.from(buffer).toString('base64');
+			return `## Media File\n**Type:** ${mimeType}\n**Size:** ${(buffer.byteLength / 1024).toFixed(1)}KB\n**Data URL:** \`data:${mimeType};base64,${base64.slice(0, 100)}...\``;
+		}
+
+		if(mimeType.match(/^text\/(plain|csv|xml)/) || args.url.match(/\.(txt|csv|xml|md|yaml|yml)$/i)) {
+			const text = await response.text();
+			const truncated = text.length > 50000 ? text.slice(0, 50000) : text;
+			return `## Text File\n**Type:** ${mimeType}\n**URL:** ${args.url}\n\n${truncated}`;
+		}
+
+		if(mimeType.match(/application\/(json|xml|csv)/)) {
+			const text = await response.text();
+			const truncated = text.length > 50000 ? text.slice(0, 50000) : text;
+			return `## Structured Data\n**Type:** ${mimeType}\n**URL:** ${args.url}\n\n\`\`\`\n${truncated}\n\`\`\``;
+		}
+
+		if(mimeType === 'application/pdf' || (mimeType.startsWith('application/') && !mimeType.includes('html'))) {
+			const buffer = await response.arrayBuffer();
+			if(buffer.byteLength > maxSize) {
+				return `❌ File too large: ${(buffer.byteLength / 1024 / 1024).toFixed(1)}MB (max 10MB)\nType: ${mimeType}`;
+			}
+			const base64 = Buffer.from(buffer).toString('base64');
+			return `## Binary File\n**Type:** ${mimeType}\n**Size:** ${(buffer.byteLength / 1024).toFixed(1)}KB\n**Data URL:** \`data:${mimeType};base64,${base64.slice(0, 100)}...\``;
+		}
+
+		// HTML
+		const html = await response.text();
+		const $ = cheerio.load(html);
+		$('script, style, nav, footer, header, aside, iframe, noscript, svg').remove();
+		$('[role="navigation"], [role="banner"], [role="complementary"]').remove();
+		$('[aria-hidden="true"], [hidden], .visually-hidden, .sr-only, .screen-reader-text').remove();
+		$('.ad, .ads, .advertisement, .cookie, .popup, .modal, .sidebar, .related, .comments, .social-share').remove();
+		$('button, [class*="share"], [class*="follow"], [class*="social"]').remove();
+		const title = $('meta[property="og:title"]').attr('content') || $('title').text().trim() || '';
+		const description = $('meta[name="description"]').attr('content') || $('meta[property="og:description"]').attr('content') || '';
+		const author = $('meta[name="author"]').attr('content') || '';
+		let content = '';
+		const selectors = ['article', 'main', '[role="main"]', '.content', '.post-content', '.entry-content', '.article-content'];
+		for(const sel of selectors) {
+			const el = $(sel).first();
+			if(el.length && el.text().trim().length > 200) {
+				const paragraphs: string[] = [];
+				el.find('p').each((_, p) => {
+					const text = $(p).text().trim();
+					if(text.length > 80) paragraphs.push(text);
+				});
+				if(paragraphs.length > 2) {
+					content = paragraphs.join('\n\n');
+					break;
+				}
+			}
+		}
+
+		if(!content) {
+			const paragraphs: string[] = [];
+			$('body p').each((_, p) => {
+				const text = $(p).text().trim();
+				if(text.length > 80) paragraphs.push(text);
+			});
+			content = paragraphs.slice(0, 30).join('\n\n');
+		}
+
+		// Decode escaped newlines and clean
+		const parts = [`## ${title || 'Webpage'}`];
+		if(description) parts.push(`_${description}_`);
+		if(author) parts.push(`👤 ${author}`);
+		parts.push(`🔗 ${args.url}\n`);
+		parts.push(content);
+		return decodeHtml(parts.join('\n\n').replaceAll(/\n{3,}/g, '\n\n'));
+	}
+};
+
+export const WebSearchTool: AiTool = {
+	name: 'web_search',
+	description: 'Use duckduckgo (anonymous) to find find relevant online resources. Returns a list of URLs that works great with the `read_webpage` tool',
 	args: {
 		query: {type: 'string', description: 'Search string', required: true},
 		length: {type: 'string', description: 'Number of results to return', default: 5},
@@ -123,7 +804,7 @@ export const SearchTool: AiTool = {
 		length: number;
 	}) => {
 		const html = await fetch(`https://html.duckduckgo.com/html/?q=${encodeURIComponent(args.query)}`, {
-			headers: {"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64)", "Accept-Language": "en-US,en;q=0.9"}
+			headers: {"User-Agent": UA, "Accept-Language": "en-US,en;q=0.9"}
 		}).then(resp => resp.text());
 		let match, regex = /<a .*?href="(.+?)".+?<\/a>/g;
 		const results = new ASet<string>();

src/vision.ts

@@ -0,0 +1,42 @@
+import {createWorker} from 'tesseract.js';
+import {AbortablePromise, Ai} from './ai.ts';
+
+export class Vision {
+
+	constructor(private ai: Ai) {}
+
+	/**
+	 * Convert image to text using Optical Character Recognition
+	 * @param {string} path Path to image
+	 * @returns {AbortablePromise<string | null>} Promise of extracted text with abort method
+	 */
+	ocr(path: string): AbortablePromise<string | null> {
+		let worker: any;
+		let reject: (err: any) => void;
+
+		const handler = (err: Error) => {
+			if(err.stack?.includes('tesseract.js')) {
+				process.off('uncaughtException', handler);
+				reject?.(err);
+				return;
+			}
+			throw err;
+		};
+		process.on('uncaughtException', handler);
+
+		const p = (async () => {
+			worker = await createWorker(this.ai.options.ocr || 'eng', 2, {cachePath: this.ai.options.path});
+			return await new Promise<string | null>((res, rej) => {
+				reject = rej;
+				worker.recognize(path)
+					.then(({data}: any) => res(data.text.trim() || null))
+					.catch(rej);
+			});
+		})().finally(() => {
+			process.off('uncaughtException', handler);
+			worker?.terminate();
+		});
+
+		return Object.assign(p, {abort: () => worker?.terminate()});
+	}
+}

tests/llm.spec.ts

@@ -0,0 +1,167 @@
+
+import {describe, it, expect, vi, beforeEach} from 'vitest';
+import LLM from '../src/llm';
+
+const {FakeProvider, providerLog} = vi.hoisted(() => {
+	const providerLog: any[] = [];
+	class FakeProvider {
+		model: string;
+		constructor(...args: any[]) { this.model = args[args.length - 1]; }
+		ask(message: string, opts: any) {
+			let aborted = false;
+			const p = (async () => {
+				const script = (globalThis as any).__scripts?.[this.model];
+				const plan = script ? script(message, opts) : {text: ''};
+				providerLog.push({model: this.model, message, system: opts.system, tools: (opts.tools || []).map((t: any) => t.name)});
+				for (const c of plan.calls || []) {
+					if (aborted) break;
+					const tool = (opts.tools || []).find((t: any) => t.name === c.tool);
+					const id = c.id || `${c.tool}_${Math.random()}`;
+					const content = await tool.fn(c.args, opts.stream, null, id);
+					opts.history.push({role: 'tool', id, name: c.tool, args: c.args, content, timestamp: Date.now()});
+				}
+				const text = plan.text ?? '';
+				if (opts.stream && text) opts.stream({text, done: true});
+				opts.history.push({role: 'assistant', content: text, timestamp: Date.now(), duration: 10, tps: 5});
+				return text;
+			})();
+			return Object.assign(p, {abort: () => { aborted = true; }});
+		}
+	}
+	return {FakeProvider, providerLog};
+});
+
+vi.mock('../src/antrhopic.ts', () => ({Anthropic: FakeProvider}));
+vi.mock('../src/open-ai.ts', () => ({OpenAi: FakeProvider}));
+
+function makeAi(models: any) {
+	return {options: {llm: {models}}} as any;
+}
+
+beforeEach(() => {
+	providerLog.length = 0;
+	(globalThis as any).__scripts = {};
+});
+
+describe('LLM cross-provider interchangeability', () => {
+	it('runs identical tool calls the same way on an anthropic-backed model and an openai-backed model', async () => {
+		const ai = makeAi({
+			claude: {proto: 'anthropic', token: 'x'},
+			gpt: {proto: 'openai', token: 'y', host: 'http://local'},
+		});
+		const llm = new LLM(ai);
+		const calc = {
+			name: 'calc_add',
+			description: 'Add two numbers',
+			args: {a: {type: 'number', required: true}, b: {type: 'number', required: true}},
+			fn: (args: any) => String(args.a + args.b),
+		};
+
+		(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'calc_add', args: {a: 2, b: 3}}], text: 'Result: 5'});
+		(globalThis as any).__scripts.gpt = () => ({calls: [{tool: 'calc_add', args: {a: 2, b: 3}}], text: 'Result: 5'});
+
+		const historyA: any[] = [], historyB: any[] = [];
+		const respA = await llm.ask('add 2 and 3', {model: 'claude', tools: [calc], history: historyA});
+		const respB = await llm.ask('add 2 and 3', {model: 'gpt', tools: [calc], history: historyB});
+
+		expect(respA).toBe('Result: 5');
+		expect(respB).toBe('Result: 5');
+		expect(providerLog.find(l => l.model === 'claude')!.tools).toContain('calc_add');
+		expect(providerLog.find(l => l.model === 'gpt')!.tools).toContain('calc_add');
+
+		// tool timing gets recomputed from real execution regardless of proto
+		for (const h of [historyA.find(h => h.name === 'calc_add'), historyB.find(h => h.name === 'calc_add')]) {
+			expect(h.content).toBe('5');
+			expect(typeof h.duration).toBe('number');
+			expect(typeof h.tps).toBe('number');
+		}
+	});
+
+	it('lets the same shared history flow across model + proto swaps with different system prompts', async () => {
+		const ai = makeAi({
+			claude: {proto: 'anthropic', token: 'x'},
+			gpt: {proto: 'openai', token: 'y', host: 'http://local'},
+		});
+		const llm = new LLM(ai);
+		const history: any[] = [];
+
+		(globalThis as any).__scripts.claude = () => ({text: 'Hi from claude'});
+		(globalThis as any).__scripts.gpt = () => ({text: 'Hi from gpt'});
+
+		const r1 = await llm.ask('hello', {model: 'claude', system: 'You are terse.', history});
+		const r2 = await llm.ask('follow up', {model: 'gpt', system: 'You are verbose.', history});
+
+		expect(r1).toBe('Hi from claude');
+		expect(r2).toBe('Hi from gpt');
+		expect(history.filter(h => h.role === 'assistant').map(h => h.content)).toEqual(['Hi from claude', 'Hi from gpt']);
+		expect(providerLog[0].system).toContain('You are terse.');
+		expect(providerLog[1].system).toContain('You are verbose.');
+	});
+
+	it('exposes MCP tools the same way no matter which proto backs the model', async () => {
+		const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
+		const llm = new LLM(ai);
+		const mcp = [{name: 'weather', host: 'http://mcp.local'}];
+
+		global.fetch = vi.fn(async (url: string, opts?: any) => {
+			if (url.endsWith('/tools')) {
+				return {json: async () => ({tools: [{name: 'lookup', description: 'Look up weather', inputSchema: {properties: {city: {type: 'string'}}, required: ['city']}}]})} as any;
+			}
+			const body = JSON.parse(opts.body);
+			return {json: async () => ({content: [{text: `Sunny in ${body.arguments.city}`}]})} as any;
+		}) as any;
+
+		for (const model of ['claude', 'gpt']) {
+			(globalThis as any).__scripts[model] = () => ({calls: [{tool: 'weather_lookup', args: {city: 'Rome'}}], text: 'done'});
+			const history: any[] = [];
+			await llm.ask('weather?', {model, mcp, history});
+			expect(history.find(h => h.name === 'weather_lookup')?.content).toBe('Sunny in Rome');
+		}
+	});
+
+	it('exposes and resolves skill documents identically across protos', async () => {
+		const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
+		const llm = new LLM(ai);
+		const skills = [{name: 'Onboarding', description: 'How to onboard a user', content: 'Step 1...'}];
+
+		for (const model of ['claude', 'gpt']) {
+			(globalThis as any).__scripts[model] = () => ({calls: [{tool: 'skill_read', args: {name: 'Onboarding'}}], text: 'done'});
+			const history: any[] = [];
+			await llm.ask('onboard me', {model, skills, history});
+			expect(history.find(h => h.name === 'skill_read')?.content).toContain('Step 1...');
+		}
+	});
+
+	it('delegate agent mutates the shared history directly and backfills the orchestrator response, across protos', async () => {
+		const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
+		const llm = new LLM(ai);
+		const history: any[] = [{role: 'user', content: 'research quantum computing'}];
+		const researcher = {name: 'researcher', system: 'You research topics.', delegate: true, model: 'gpt'};
+
+		(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'agent_researcher', args: {}}], text: ''});
+		(globalThis as any).__scripts.gpt = () => ({text: 'Quantum computers use qubits.'});
+
+		const resp = await llm.ask('go', {model: 'claude', agents: [researcher], history});
+
+		expect(resp).toBe('Quantum computers use qubits.');
+		expect(history.some(h => h.role === 'assistant' && h.content === 'Quantum computers use qubits.')).toBe(true);
+		expect(history.find(h => h.name === 'agent_researcher')?.content).toBe('');
+	});
+
+	it('regular (non-delegate) subagent keeps its own isolated history separate from the parent, across protos', async () => {
+		const ai = makeAi({claude: {proto: 'anthropic', token: 'x'}, gpt: {proto: 'openai', token: 'y', host: 'http://local'}});
+		const llm = new LLM(ai);
+		const history: any[] = [];
+		const summarizer = {name: 'summarizer', system: 'You summarize text.', model: 'gpt'};
+
+		(globalThis as any).__scripts.claude = () => ({calls: [{tool: 'subagent_summarizer', args: {context: 'a long article', instructions: 'summarize it'}}], text: 'Summary: short version'});
+		(globalThis as any).__scripts.gpt = () => ({text: 'short version'});
+
+		const resp = await llm.ask('summarize this', {model: 'claude', agents: [summarizer], history});
+
+		expect(resp).toBe('Summary: short version');
+		expect(history.find(h => h.name === 'subagent_summarizer')?.content).toBe('short version');
+		// isolated history - subagent's own assistant turn never leaks into the parent
+		expect(history.some(h => h.role === 'assistant' && h.content === 'short version')).toBe(false);
+	});
+});

tests/memory.spec.ts

@@ -0,0 +1,256 @@
+import {describe, it, expect, vi, beforeEach} from 'vitest';
+import {MemoryManager, MemoryCache, rebuildGraph, Memory} from '../src/memory';
+
+function makeMemory(overrides: Partial<Memory> = {}): Memory {
+	return {
+		name: 'Test/Doc',
+		description: '',
+		content: '',
+		embedding: [],
+		links: [],
+		backlinks: [],
+		...overrides,
+	};
+}
+
+function makeLLM() {
+	return {
+		embedding: vi.fn(async (_text: string) => [{embedding: [1, 0, 0]}]),
+		ask: vi.fn(async () => undefined),
+	};
+}
+
+describe('rebuildGraph', () => {
+	it('extracts [[WikiLinks]] from content, excluding self-links', () => {
+		const a = makeMemory({name: 'A', content: '[[B]] and [[A]] and [[C]]'});
+		const b = makeMemory({name: 'B', content: 'no links here'});
+		const mem = [a, b];
+
+		rebuildGraph(mem);
+
+		expect(a.links).toEqual(['B', 'C']);
+		expect(b.links).toEqual([]);
+	});
+
+	it('computes backlinks only for links that resolve to a real node', () => {
+		const a = makeMemory({name: 'A', content: '[[B]] [[Missing]]'});
+		const b = makeMemory({name: 'B', content: ''});
+		const mem = [a, b];
+
+		rebuildGraph(mem);
+
+		expect(b.backlinks).toEqual(['A']);
+		expect(mem.find(m => m.name === 'Missing')).toBeUndefined();
+	});
+
+	it('resets stale backlinks on every rebuild (no leftover from a removed link)', () => {
+		const a = makeMemory({name: 'A', content: '[[B]]'});
+		const b = makeMemory({name: 'B', content: ''});
+		const mem = [a, b];
+		rebuildGraph(mem);
+		expect(b.backlinks).toEqual(['A']);
+
+		a.content = 'no more links';
+		rebuildGraph(mem);
+		expect(b.backlinks).toEqual([]);
+	});
+});
+
+describe('MemoryCache', () => {
+	it('finds nearest neighbor by embedding via KD-tree search', () => {
+		const close = makeMemory({name: 'Close', embedding: [1, 0, 0]});
+		const far = makeMemory({name: 'Far', embedding: [0, 0, 1]});
+		const cache = new MemoryCache([close, far]);
+
+		const results = cache.search([1, 0, 0], 1);
+
+		expect(results[0].name).toBe('Close');
+	});
+
+	it('rebuilds the tree on add/update/remove', () => {
+		const cache = new MemoryCache([makeMemory({name: 'A', embedding: [1, 0, 0]})]);
+		cache.add(makeMemory({name: 'B', embedding: [0, 1, 0]}));
+		expect(cache.search([0, 1, 0], 1)[0].name).toBe('B');
+
+		cache.remove('B');
+		expect(cache.search([0, 1, 0], 1)[0]?.name).not.toBe('B');
+	});
+});
+
+describe('MemoryManager.forget', () => {
+	it('removes the node and recomputes backlinks for the rest of the graph', () => {
+		const llm = makeLLM();
+		const mgr = new MemoryManager(llm);
+		const a = makeMemory({name: 'A', content: '[[B]]'});
+		const b = makeMemory({name: 'B', content: '[[C]]'});
+		const c = makeMemory({name: 'C', content: ''});
+		const mem = [a, b, c];
+		rebuildGraph(mem);
+		expect(c.backlinks).toEqual(['B']);
+
+		const ok = mgr.forget('B', mem);
+
+		expect(ok).toBe(true);
+		expect(mem.find(m => m.name === 'B')).toBeUndefined();
+		expect(a.links).toEqual(['B']);
+		expect(c.backlinks).toEqual([]);
+	});
+
+	it('returns false for an unknown name', () => {
+		const mgr = new MemoryManager(makeLLM());
+		expect(mgr.forget('Nope', [makeMemory({name: 'A'})])).toBe(false);
+	});
+});
+
+describe('MemoryManager.recollect', () => {
+	it('orders vector matches first, then expands one hop via links', async () => {
+		const llm = makeLLM();
+		llm.embedding.mockResolvedValue([{embedding: [1, 0, 0]}]);
+		const mgr = new MemoryManager(llm);
+
+		const near = makeMemory({name: 'Near', embedding: [1, 0, 0], content: '[[Linked]]'});
+		const linked = makeMemory({name: 'Linked', embedding: [0, 0, 1], content: ''});
+		const far = makeMemory({name: 'Far', embedding: [0, 1, 0], content: ''});
+		const mem = [near, linked, far];
+		rebuildGraph(mem);
+
+		const result = await mgr.recollect('query', mem, 1, 1);
+
+		expect(result.map(r => r.name)).toEqual(['Near', 'Linked']);
+	});
+
+	it('returns [] when there are no memories', async () => {
+		const mgr = new MemoryManager(makeLLM());
+		expect(await mgr.recollect('q', [])).toEqual([]);
+	});
+});
+
+describe('MemoryManager.memorize (fast path)', () => {
+	let llm: ReturnType<typeof makeLLM>;
+	let mgr: MemoryManager;
+
+	beforeEach(() => {
+		llm = makeLLM();
+		mgr = new MemoryManager(llm);
+	});
+
+	it('pushes a pending tool message, then resolves it to links once facts land', async () => {
+		llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
+			if (opts.tools) {
+				opts.tools[0].fn({destination: 'Projects/Oxide', facts: 'Uses a hybrid memory system'});
+				return undefined;
+			}
+			return {description: 'd', content: '# doc'};
+		});
+
+		const history: any[] = [{role: 'user', content: 'we use a hybrid memory system'}];
+		const touched = await mgr.memorize(history, [], {model: 'test'} as any);
+
+		const pending = history.find(h => h.name === 'memory_process');
+		expect(pending).toBeDefined();
+		expect(pending.content).toContain('[[Projects/Oxide]]');
+		expect(touched.map(t => t.name)).toEqual(['Projects/Oxide']);
+	});
+
+	it('creates a new node and appends facts under "## Facts" without calling the doc LLM', async () => {
+		llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
+			if (opts.tools) opts.tools[0].fn({destination: 'People/Sarah', facts: 'Works at Acme, Likes hiking'});
+			return undefined;
+		});
+
+		const mem: Memory[] = [];
+		await mgr.memorize([{role: 'user', content: 'Sarah works at Acme and likes hiking'}] as any, mem, {model: 'test'} as any);
+
+		const node = mem.find(m => m.name === 'People/Sarah')!;
+		expect(node).toBeDefined();
+		expect(node.content).toContain('## Facts');
+		expect(node.content).toContain('- Works at Acme');
+		expect(node.content).toContain('- Likes hiking');
+		// doc reconciler LLM (schema call) should NOT have been awaited synchronously in this fast path assertion
+	});
+
+	it('routes "journal" destination to Journal/{weekMonday}', async () => {
+		llm.ask.mockImplementation(async (_prompt: string, opts: any) => {
+			if (opts.tools) opts.tools[0].fn({destination: 'journal', facts: 'Shipped v1'});
+			return undefined;
+		});
+
+		const mem: Memory[] = [];
+		const touched = await mgr.memorize([{role: 'user', content: 'shipped v1 today'}] as any, mem, {model: 'test'} as any);
+
+		expect(touched[0].name).toMatch(/^Journal\/\d{4}-\d{2}-\d{2}$/);
+	});
+
+	it('reports nothing to remember when no facts are extracted', async () => {
+		llm.ask.mockResolvedValue(undefined); // tools present but fn never called
+
+		const history: any[] = [{role: 'user', content: 'hey'}];
+		const touched = await mgr.memorize(history, [], {model: 'test'} as any);
+
+		expect(touched).toEqual([]);
+		expect(history.find(h => h.name === 'memory_process').content).toBe('Nothing worth remembering.');
+	});
+
+	it('returns [] and does nothing for an empty conversation', async () => {
+		const touched = await mgr.memorize([], [], {model: 'test'} as any);
+		expect(touched).toEqual([]);
+		expect(llm.ask).not.toHaveBeenCalled();
+	});
+});
+
+describe('MemoryManager reconcileVault', () => {
+	it('integrates the "## Facts" section via the doc LLM and removes it', async () => {
+		const llm = makeLLM();
+		llm.ask.mockResolvedValue({description: 'Tidy summary', content: '# Doc\n\nIntegrated fact.'});
+		const mgr = new MemoryManager(llm);
+
+		const node = makeMemory({
+			name: 'Projects/Oxide',
+			content: '---\nname: Projects/Oxide\n---\n\n# Doc\n\n## Facts\n- some raw fact\n',
+		});
+		const mem = [node];
+
+		await mgr.reconcileVault(mem, {model: 'test'} as any, 'all');
+
+		expect(node.content).not.toContain('## Facts');
+		expect(node.content).toContain('Integrated fact.');
+		expect(node.description).toBe('Tidy summary');
+	});
+
+	it('only targets docs with a pending Facts inbox when scope is "touched"', async () => {
+		const llm = makeLLM();
+		llm.ask.mockResolvedValue({description: 'd', content: '# clean'});
+		const mgr = new MemoryManager(llm);
+
+		const dirty = makeMemory({name: 'A', content: '## Facts\n- x'});
+		const clean = makeMemory({name: 'B', content: '# already tidy'});
+		await mgr.reconcileVault([dirty, clean], {model: 'test'} as any, 'touched');
+
+		expect(dirty.content).toContain('# clean'); // rewritten (frontmatter now wraps it)
+		expect(clean.content).toBe('# already tidy'); // untouched, never queued
+	});
+});
+
+describe('MemoryManager reconcile coalescing', () => {
+	it('coalesces a second call while one is in-flight: marks dirty, aborts, reuses the same task promise', () => {
+		const llm = makeLLM();
+		const abort = vi.fn();
+		let calls = 0;
+		llm.ask.mockImplementation(() => {
+			calls++;
+			const pending: any = new Promise(() => {}); // never resolves in this test
+			pending.abort = abort;
+			return pending;
+		});
+		const mgr: any = new MemoryManager(llm);
+		const node = makeMemory({name: 'Q', content: '# Q\n\n## Facts\n- f'});
+		const mem = [node];
+
+		const p1 = mgr.reconcile(node, mem, {model: 'test'});
+		const p2 = mgr.reconcile(node, mem, {model: 'test'});
+
+		expect(p2).toBe(p1); // same in-flight task, not a new queue entry
+		expect(abort).toHaveBeenCalledTimes(1); // second call aborted the in-flight request
+		expect(calls).toBe(1); // no second ask() fired synchronously — it'll rerun via the dirty loop
+	});
+});

tsconfig.json

@@ -4,7 +4,10 @@
 		"target": "ESNext",
 		"useDefineForClassFields": true,
 		"module": "ESNext",
-		"lib": ["ESNext"],
+		"lib": [
+			"ESNext",
+			"dom"
+		],
 		"skipLibCheck": true,
 
 		/* Bundler mode */
@@ -15,6 +18,7 @@
 		"noEmit": true,
 
 		/* Linting */
-		"strict": true
+		"strict": true,
+		"noImplicitAny": false
 	}
 }

vite.config.ts

@@ -1,20 +1,22 @@
-import {resolve} from 'path';
 import {defineConfig} from 'vite';
 import dts from 'vite-plugin-dts';
 
 export default defineConfig({
 	build: {
 		lib: {
-			entry: resolve(process.cwd(), 'src/index.ts'),
+			entry: {
+				index: './src/index.ts',
+				embedder: './src/embedder.ts',
+			},
 			name: 'utils',
-			fileName: (module, entryName) => {
+			fileName: (format, entryName) => {
-				if(module == 'es') return 'index.mjs';
+				if (entryName === 'embedder') return 'embedder.js';
-				if(module == 'umd') return 'index.cjs';
+				return format === 'es' ? 'index.mjs' : 'index.js';
-			}
+			},
 		},
 		ssr: true,
 		emptyOutDir: true,
-		minify: false,
+		minify: true,
 		sourcemap: true
 	},
 	plugins: [dts()],