@fkws/klonk 0.0.4 → 0.0.6

package/README.md

@@ -1,276 +1,449 @@
-<picture>
-  <source media="(prefers-color-scheme: dark)" srcset=".github/assets/logo_dark_mode.png">
-  <source media="(prefers-color-scheme: light)" srcset=".github/assets/logo_bright_mode.png">
-  <img alt="Klonk Logo" src=".github/assets/logo_bright_mode.png">
-</picture>
-
 # Klonk
+*A code-first, type-safe automation engine for TypeScript.*
 
-Klonk is a **code-first, self-hosted automation engine** for TypeScript developers who demand true end-to-end type safety. It enables you to build complex, event-driven workflows with a fluent, declarative API.
+## Introduction
+Klonk is a code-first, type-safe automation engine designed with developer experience as a top priority. It provides powerful, composable primitives to build complex workflows and state machines with world-class autocomplete and type inference. If you've ever wanted to build event-driven automations or a stateful agent, but in code, with all the benefits of TypeScript, Klonk is for you.
 
-Stop wrestling with `any` types and stringly-typed data blobs between your automation steps. With Klonk, your entire workflow is a single, statically-checked TypeScript application, from the initial trigger to the final task.
+The two main features are **Workflows** and **Machines**.
 
-## Core Principles
+- **Workflows**: Combine triggers with a series of tasks (a `Playlist`) to automate processes. Perfect for event-driven automation, like "when a file is added to Dropbox, parse it, and create an entry in Notion."
+- **Machines**: Create finite state machines where each state has its own `Playlist` of tasks and conditional transitions to other states. Ideal for building agents, multi-step processes, or any system with complex, stateful logic.
 
-Klonk is designed to provide a developer experience focused on type safety and a code-first approach.
+## Installation
+```bash
+bun add @fkws/klonk
+# or
+npm i @fkws/klonk
+```
 
-### End-to-End Type Safety
+## Core Concepts
 
-Klonk's fluent API leverages TypeScript's type system to provide compile-time safety across your entire workflow. As you add tasks to a `Playlist`, the `outputs` object available to subsequent tasks is **automatically and cumulatively typed**.
+At the heart of Klonk are a few key concepts that work together.
 
-```typescript
-.setPlaylist(p => p
-  .addTask(new A_Task("task-a", client), /* ... */)
-  // In the next step, `outputs["task-a"]` is fully typed! No `any`, no manual casting.
-  .addTask(new B_Task("task-b", client), (source, outputs) => ({
-      inputForB: outputs['task-a'].someProperty // Easily leverage auto-completion in your IDE
-  }))
-);
-```
+### Task
+A `Task` is the smallest unit of work. It's an abstract class with two main methods you need to implement:
+- `validateInput(input)`: Runtime validation of the task's input (on top of strong typing).
+- `run(input)`: Executes the task's logic.
 
-This means you get **flawless autocompletion** and can **prevent runtime errors** before your code ever runs, providing a level of safety and productivity not found in many GUI-based tools or other code-based orchestrators.
+Tasks use a `Railroad` return type, which is a simple way to handle success and error states without throwing exceptions. You can also use it for the rest of your application.
 
-### A Code-First Approach
+### Playlist
+A `Playlist` is a sequence of `Tasks` that are executed in order. The magic of a `Playlist` is that each task has access to the outputs of all the tasks that ran before it, in a fully type-safe way. You build a `Playlist` by chaining `.addTask()` calls.
 
-Klonk is built for developers who prefer to work with code. You use the full power and expressiveness of TypeScript to:
--   Implement complex logic and data transformations with ease.
--   Version control your workflows with Git.
--   Write unit and integration tests.
--   Integrate with your existing CI/CD pipelines.
+### Trigger
+A `Trigger` is what kicks off a `Workflow`. It's an event source. Klonk can be extended with triggers for anything: file system events, webhooks, new database entries, messages in a queue, etc.
 
-### Lightweight & Simple to Self-Host
+### Workflow
+A `Workflow` connects one or more `Triggers` to a `Playlist`. When a trigger fires an event, the workflow runs the playlist, passing the event data as the initial input. This allows you to create powerful, event-driven automations.
 
-Klonk is incredibly lightweight, running as a simple Node.js or Bun process with no complex dependencies or external infrastructure. This makes it a simpler alternative to heavy-duty orchestrators like Airflow or Temporal. You can get a workflow up and running in minutes on your own server with no subscription fees.
+### Machine
+A `Machine` is a finite state machine. It's made up of `StateNode`s. Each `StateNode` represents a state and has two key components:
+1.  A `Playlist` that runs when the machine enters that state.
+2.  A set of conditional `Transitions` to other states.
+3.  Retry rules for when a transition fails to resolve.
 
-### Powerful Built-Ins
-`@fkws-npm/klonk/tasks` and `@fkws-npm/klonk/triggers` implement powerful built-ins. Klonk provides a rich set of triggers and tasks that cover a wide range of common automation scenarios, from file system events to webhook listeners. These components are designed to be robust and ready for production use. When combined with the official integrations for popular services like Dropbox, Notion, and OpenRouter, you have all the tools you need to build versatile and powerful workflows right out of the box, without having to write custom components for every step.
+The `Machine` carries a mutable `stateData` object that can be read from and written to by playlists and transition conditions throughout its execution.
 
 ## Features
+- **Type-Safe & Autocompleted**: Klonk leverages TypeScript's inference to provide a world-class developer experience. The inputs and outputs of every step are strongly typed, so you'll know at compile time if your logic is sound.
+- **Code-First**: Define your automations directly in TypeScript. No YAML, no drag-and-drop UIs. Just the full power of a real programming language.
+- **Composable & Extensible**: The core primitives (`Task`, `Trigger`) are simple abstract classes, making it easy to create your own reusable components and integrations.
+- **Flexible Execution**: `Machines` can be run synchronously to completion (`run`) for request/response style work, or started as a long-running background process (`start`).
 
-- **Statically-Checked Workflows**: Build complex automations with a fluent API where data passed between steps is fully type-safe.
-- **Modular by Design**: Easily extend the engine with your own custom triggers and tasks.
-- **Modern Integration Support**: Built-in, async-ready integrations for services like Notion, Dropbox, and OpenRouter.
-- **AI-Powered**: Leverage powerful AI models for intelligent document and data processing right out of the box.
-- **Code-First & Declarative**: Define workflows in pure TypeScript for maximum flexibility and clarity.
-- **Self-Hosted**: Run on your own servers with no recurring fees.
-
-## Installation
+## Klonkworks: Pre-built Components
+Coming soon(ish)! Klonkworks will be a large collection of pre-built Tasks, Triggers, and integrations. This will allow you to quickly assemble powerful automations that connect to a wide variety of services, often without needing to build your own components from scratch.
 
-```bash
-bun add @fkws-npm/klonk
-```
+## Code Examples
+<details>
+<summary><b>Creating a Task</b></summary>
 
-Or with npm:
+Here's how you create a custom `Task`. This task uses an AI client to perform text inference.
 
-```bash
-npm install @fkws-npm/klonk
+```typescript
+import { Railroad, Task } from "@fkws/klonk";
+import { OpenRouterClient } from "./common/OpenrouterClient"
+import { Model } from "./common/models";
+
+type TABasicTextInferenceInput = {
+    inputText: string;
+    instructions?: string;
+    model: Model;
+};
+
+type TABasicTextInferenceOutput = {
+    text: string;
+};
+
+// A Task is a generic class. You provide the Input, Output, and an Ident (a unique string literal for the task).
+export class TABasicTextInference<IdentType extends string> extends Task<
+    TABasicTextInferenceInput,  // These type parameters are part of the secret sauce typing system Klonk uses.
+    TABasicTextInferenceOutput, // Input Type, Output Type, Ident Type
+    IdentType
+> {
+    constructor(ident: IdentType, public client: OpenRouterClient) {
+        super(ident);
+        if (!this.client) {
+            throw new Error("[TABasicTextInference] An IOpenRouter client instance is required.");
+        }
+    }
+
+    // validateInput is for runtime validation of the data your task receives.
+    async validateInput(input: TABasicTextInferenceInput): Promise<boolean> {
+        if (!input.inputText || !input.model) {
+            return false;
+        }
+        return true;
+    }
+
+    // The core logic of your task. It must return a Railroad type.
+    async run(input: TABasicTextInferenceInput): Promise<Railroad<TABasicTextInferenceOutput>> {
+        try {
+            const result = await this.client.basicTextInference({
+                inputText: input.inputText,
+                instructions: input.instructions,
+                model: input.model
+            });
+            // On success, return a success object with your data.
+            return {
+                success: true, // Railroad is a simple result type
+                data: {
+                    text: result
+                }
+            };
+        } catch (error) {
+            // On failure, return an error object. The next Task's input builder will react to this.
+            return {
+                success: false,
+                error: error instanceof Error ? error : new Error(String(error))
+            };
+        }
+    }
+}
 ```
+</details>
 
-## Core Concepts
+<details>
+<summary><b>Creating a Trigger</b></summary>
 
-Klonk is built around a few key concepts that work together to create powerful automations:
+Here's an example of a custom `Trigger`. This trigger fires on a given interval and pushes the current date as its event data.
 
-- **Workflow**: The main orchestrator. You build a workflow by adding triggers and then defining a single playlist of tasks to execute.
-- **Triggers**: Event sources that start a workflow's playlist. A trigger could be a file being added to a Dropbox folder, a webhook being called, or a scheduled timer. Each time a trigger fires, it produces an event that is passed to the playlist.
-- **Playlist**: An ordered sequence of tasks. The playlist is run once for every event produced by a trigger. It intelligently carries the typed outputs of each task to the ones that follow.
-- **Tasks**: The individual, atomic actions that make up a playlist, such as calling an API, downloading a file, or processing data. The output of one task is available by its unique ID to all subsequent tasks.
+```typescript
+import { Trigger } from '@fkws/klonk';
+
+// A simple trigger that fires every `intervalMs` with the current date.
+// You define the shape of the data the trigger will provide, in this case `{ now: Date }`.
+export class IntervalTrigger<TIdent extends string> extends Trigger<TIdent, { now: Date }> {
+    private intervalId: NodeJS.Timeout | null = null;
+
+    constructor(ident: TIdent, private intervalMs: number) {
+        super(ident); // Pass the unique identifier to the parent constructor.
+    }
+
+    // The start method is called by the Workflow to begin listening for events.
+    async start(): Promise<void> {
+        if (this.intervalId) return; // Prevent multiple intervals.
+
+        this.intervalId = setInterval(() => {
+            // When an event occurs, use pushEvent to add it to the internal queue for the workflow to poll.
+            this.pushEvent({ now: new Date() });
+        }, this.intervalMs);
+    }
+
+    // The stop method cleans up any resources, like intervals or open connections.
+    async stop(): Promise<void> {
+        if (this.intervalId) {
+            clearInterval(this.intervalId);
+            this.intervalId = null;
+        }
+    }
+}
+```
+</details>
 
-## Getting Started: A Simple Workflow
+<details>
+<summary><b>Building a Workflow</b></summary>
 
-Here's a basic example that demonstrates the core concepts. This workflow triggers when a file is added to a Dropbox folder, downloads it, and then logs its contents.
+Workflows are perfect for event-driven automations. This example creates a workflow that triggers when a new invoice PDF is added to a Dropbox folder. It then parses the invoice and creates a new item in a Notion database.
 
-```typescript
-import { Workflow } from '@fkws-npm/klonk';
-import { TRDropboxFileAdded } from '@fkws-npm/klonk/triggers';
-import { TADropboxDownloadFile, TALogToConsole } from '@fkws-npm/klonk/tasks';
-import { IDropbox } from '@fkws-npm/klonk/integrations';
+Notice how the `builder` function for each task (`(source, outputs) => { ... }`) has access to the initial `source` data (from the trigger) and the `outputs` of all previous tasks. Klonk automatically infers the types for `source` and `outputs`!
 
-// 1. Initialize your service clients
-const dropboxClient = new IDropbox({
+```typescript
+import { z } from 'zod';
+import { Workflow } from '@fkws/klonk';
+
+// The following example requires a lot of tasks, integrations and a trigger.
+// Soon, you will be able to import these from @fkws/klonkworks.
+import { TACreateNotionDatabaseItem, TANotionGetTitlesAndIdsForDatabase, TAParsePdfAi, TADropboxDownloadFile } from '@fkws/klonkworks/tasks';
+import { INotion, IOpenRouter, IDropbox } from '@fkws/klonkworks/integrations';
+import { TRDropboxFileAdded } from '@fkws/klonkworks/triggers';
+
+// Providers and clients are instantiated as usual.
+const notionProvider = new INotion({apiKey: process.env.NOTION_API_KEY!});
+const openrouterProvider = new IOpenRouter({apiKey: process.env.OPENROUTER_API_KEY!});
+const dropboxProvider = new IDropbox({
     appKey: process.env.DROPBOX_APP_KEY!,
     appSecret: process.env.DROPBOX_APP_SECRET!,
-    refreshToken: process.env.DROPBOX_REFRESH_TOKEN!,
+    refreshToken: process.env.DROPBOX_REFRESH_KEY!
 });
 
-// 2. Define the workflow using the fluent API
-const simpleWorkflow = Workflow.create()
-  // Add a trigger. The type of `source.data` in the playlist will be inferred from this.
-  .addTrigger(
-    new TRDropboxFileAdded("new-file-in-dropbox", {
-      client: dropboxClient,
-      folderPath: "/MyKlonkFiles"
+// Start building a workflow.
+const workflow = Workflow.create().addTrigger(
+    // A workflow is initiated by one or more triggers.
+    new TRDropboxFileAdded("dropbox-trigger", {
+        client: dropboxProvider,
+        folderPath: process.env.DROPBOX_INVOICES_FOLDER_PATH ?? "",
     })
-  )
-  // Define the sequence of tasks to run when the trigger fires.
-  .setPlaylist(p =>
-    p.addTask(
-      // Each task has a unique identifier.
-      new TADropboxDownloadFile("download-the-file", dropboxClient),
-      // The builder function defines the task's input.
-      // `source` is the event from the trigger. `outputs` contains results from previous tasks.
-      (source, outputs) => ({
-        file_metadata: source.data // Fully typed, autocomplete-ready and will yell if wrong.
-      })
-    ) // -> Produces { file: Buffer }
-    .addTask(
-      // The second task logs the content of the downloaded file.
-      new TALogToConsole("log-the-content"),
-      (source, outputs) => ({
-        // The `file` property from the previous task's output is accessed in a type-safe way so you can safely call `.toString()`.
-        message: outputs["download-the-file"].file.toString(),
-      })
+).setPlaylist(p => p // Builder function allows complex types to be assembled!
+    .addTask( // .addTask() adds a task to the playlist.
+        new TANotionGetTitlesAndIdsForDatabase("get-payees", notionProvider),
+        // The second argument to addTask builds the input for that task.
+        // `source` is the data from the trigger, `outputs` contains all previous task outputs.
+        (source, outputs) => {
+            return { database_id: process.env.NOTION_PAYEES_DATABASE_ID!}
+        }
+    ).addTask(
+        new TANotionGetTitlesAndIdsForDatabase("get-expense-types", notionProvider),
+        (source, outputs) => { // Type inference works for source and outputs!
+            return { database_id: process.env.NOTION_EXPENSE_TYPES_DATABASE_ID!}
+        }
+    ).addTask(
+        new TADropboxDownloadFile("download-invoice-pdf", dropboxProvider),
+        (source, outputs) => {
+            // The `source` object contains the trigger ident, so you can handle multiple triggers.
+            if (source.triggerIdent == "dropbox-trigger") {
+                return { file_metadata: source.data}
+            } else {
+                throw new Error(`Trigger ${source.triggerIdent} is not implemented for task download-invoice-pdf.`)
+            }
+        }
+    ).addTask(
+        new TAParsePdfAi("parse-invoice", openrouterProvider),
+        (source, outputs) => {
+            // Access the outputs of previous tasks via the `outputs` object.
+            // The keys are the idents you provided to the tasks.
+            const downloadResult = outputs['download-invoice-pdf'];
+            if (!downloadResult.success) {
+                throw downloadResult.error ?? new Error('Failed to download invoice PDF');
+            }
+
+            const payeesResult = outputs['get-payees'];
+            if (!payeesResult.success) {
+                throw payeesResult.error ?? new Error('Failed to load payees');
+            }
+
+            const expenseTypesResult = outputs['get-expense-types'];
+            if (!expenseTypesResult.success) {
+                throw expenseTypesResult.error ?? new Error('Failed to load expense types');
+            }
+
+            const payees = payeesResult.data;
+            const expenseTypes = expenseTypesResult.data;
+
+            return {
+                pdf: downloadResult.data.file,
+                instructions: "Extract data from the invoice",
+                schema: z.object({
+                    payee: z.enum(payees.map(p => p.id) as [string, ...string[]])
+                        .describe("The payee id of the invoice according to this map: " + JSON.stringify(payees, null, 2)),
+                    total: z.number()
+                        .describe("The total amount of the invoice."),
+                    invoice_date: z.string()
+                        .regex(/^\d{4}-\d{2}-\d{2}$/)
+                        .describe("The date of the invoice as an ISO 8601 string (YYYY-MM-DD)."),
+                    expense_type: z.enum(expenseTypes.map(e => e.id) as [string, ...string[]])
+                        .describe("The expense type id of the invoice according to this map: " + JSON.stringify(expenseTypes, null, 2))
+                })
+            }
+        }
+    ).addTask(
+        new TACreateNotionDatabaseItem("create-notion-invoice", notionProvider),
+        (source, outputs) => {
+            const invoiceResult = outputs['parse-invoice'];
+            if (!invoiceResult.success) {
+                throw invoiceResult.error ?? new Error('Failed to parse invoice');
+            }
+            const invoiceData = invoiceResult.data;
+            const properties = {
+                'Name': { 'title': [{ 'text': { 'content': 'Invoice' } }] },
+                'Payee': { 'relation': [{ 'id': invoiceData.payee }] },
+                'Total': { 'number': invoiceData.total },
+                'Invoice Date': { 'date': { 'start': invoiceData.invoice_date } },
+                'Expense Type': { 'relation': [{ 'id': invoiceData.expense_type }] }
+            };
+            return {
+                database_id: process.env.NOTION_INVOICES_DATABASE_ID!,
+                properties: properties
+            }
+        }
     )
-  );
-
-// 3. Start the workflow
-simpleWorkflow.start();
-console.log("Simple workflow started! Waiting for new files in Dropbox...");
-```
-
-## Advanced Example: AI-Powered Invoice Processing
-
-This example shows a real-world workflow that:
-1. Triggers when a new invoice PDF is added to a Dropbox folder.
-2. Fetches the PDF content.
-3. Uses an AI model to extract structured data from the invoice.
-4. Creates a new record in a Notion database with the extracted data.
-
-```typescript
-import { Workflow } from '@fkws-npm/klonk';
-import { z } from 'zod';
-import { TRDropboxFileAdded } from '@fkws-npm/klonk/triggers';
-import {
-  TADropboxDownloadFile,
-  TAParsePdfAi,
-  TACreateNotionDatabaseItem
-} from '@fkws-npm/klonk/tasks';
-import { IDropbox, IOpenRouter, INotion } from '@fkws-npm/klonk/integrations';
-
-// --- Client & Schema Setup ---
-const dropbox = new IDropbox({ /* ... */ });
-const openRouter = new IOpenRouter({ apiKeyEnvVar: "OPENROUTER_API_KEY" });
-const notion = new INotion({ apiKeyEnvVar: "NOTION_API_KEY" });
-
-const InvoiceSchema = z.object({
-  invoice_number: z.string(),
-  total_amount: z.number(),
-  is_paid: z.boolean(),
+)
+
+// Run the workflow
+console.log('[WCreateNotionInvoiceFromFile] Starting workflow...');
+// .start() begins the workflow's trigger polling loop.
+workflow.start({
+    // The callback is executed every time the playlist successfully completes.
+    callback: (source, outputs) => {
+        console.log('[WCreateNotionInvoiceFromFile] Workflow completed');
+        console.dir({
+            source,
+            outputs
+        }, { depth: null });
+    }
 });
-
-// --- Workflow Definition ---
-const invoiceWorkflow = Workflow.create()
-  .addTrigger(new TRDropboxFileAdded("new-invoice", { client: dropbox, folderPath: "/Invoices" }))
-  .setPlaylist(p => p
-    // Step 1: Download the file from Dropbox
-    .addTask(new TADropboxDownloadFile("download-pdf", dropbox),
-      (source, outputs) => ({
-        file_metadata: source.data
-      })
-    )
-    // Step 2: Use AI to parse the downloaded PDF
-    .addTask(new TAParsePdfAi("parse-with-ai", openRouter),
-      (source, outputs) => ({
-        pdf: outputs['download-pdf'].file, // Type-safe access to previous output by its ID
-        instructions: "Extract the invoice number, total amount, and payment status from the PDF.",
-        zodSchema: InvoiceSchema,
-      })
-    )
-    // Step 3: Create a new item in a Notion database
-    .addTask(new TACreateNotionDatabaseItem("create-notion-record", notion),
-      (source, outputs) => {
-        const invoiceData = outputs['parse-with-ai']; // Type-safe access to the parsed data
-        return {
-          database_id: 'YOUR_INVOICES_DATABASE_ID',
-          properties: {
-            'Invoice Number': { title: [{ text: { content: invoiceData.invoice_number } }] },
-            'Total': { number: invoiceData.total_amount },
-            'Status': { select: { name: invoiceData.is_paid ? "Paid" : "Unpaid" } }
-          }
-        };
-      }
-    )
-  );
-
-// --- Start the workflow ---
-invoiceWorkflow.start();
 ```
+</details>
 
-## Creating Custom Components
-
-If your use case can't be served with Klonk's built-ins, you can easily extend Klonk by implementing the `Task` and `Trigger` abstract classes.
+<details>
+<summary><b>Building a Machine</b></summary>
 
-### Custom Task Example
+`Machines` are ideal for building complex, stateful agents. This example shows a simple AI agent that takes a user's query, refines it, performs a web search, and then generates a final response.
 
-A task must define its Input and Output types and implement `validateInput` and `run` methods.
+The `Machine` manages a `StateData` object. Each `StateNode`'s `Playlist` can modify this state, and the `Transitions` between states can use it to decide which state to move to next.
 
 ```typescript
-import { Task } from '@fkws-npm/klonk';
-import type { IEmailService } from './my-email-service'; // Your custom service
-
-// 1. Define the Input and Output types
-type EmailSenderInput = { to: string; subject: string; body: string; };
-type EmailSenderOutput = { messageId: string; };
-
-// 2. Create the class, extending the generic Task
-export class TAEmailSender<IdentType extends string> extends Task<
-  EmailSenderInput,
-  EmailSenderOutput,
-  IdentType
-> {
-  // 3. The constructor receives its ident and any clients it needs
-  constructor(ident: IdentType, private emailClient: IEmailService) {
-    super(ident);
-  }
-
-  // 4. For runtime validation
-  validateInput(input: EmailSenderInput): boolean {
-    return !!input.to && !!input.subject && !!input.body;
-  }
-
-  async run(input: EmailSenderInput): Promise<EmailSenderOutput> {
-    const result = await this.emailClient.send(input);
-    return { messageId: result.id };
-  }
+import { Machine, StateNode } from "@fkws/klonk"
+import { OpenRouterClient } from "./tasks/common/OpenrouterClient" 
+import { Model } from "./tasks/common/models"
+import { TABasicTextInference } from "./tasks/TABasicTextInference"
+import { TASearchOnline } from "./tasks/TASearchOnline"
+
+type StateData = {
+    input: string;
+    output?: string;
+    model?: Model;
+    refinedInput?: string;
+    searchTerm?: string;
+    searchResults?: {
+        results: {
+            url: string;
+            title: string;
+            content: string;
+            raw_content?: string | undefined;
+            score: string;
+        }[];
+        query: string;
+        answer?: string | undefined;
+        images?: string[] | undefined;
+        follow_up_questions?: string[] | undefined;
+        response_time: string;
+    },
+    finalResponse?: string;
 }
-```
 
-### Custom Trigger Example
+const client = new OpenRouterClient(process.env.OPENROUTER_API_KEY!)
+
+const webSearchAgent = Machine
+    .create<StateData>()
+    .addState(StateNode
+        .create<StateData>()
+        .setIdent("refine_and_extract")
+        .setPlaylist(p => p // Builder function allows complex types to be assembled!
+            .addTask(new TABasicTextInference("refine", client),
+                (state, outputs) => { // This function constructs the INPUT of the task from the state and outputs of previous tasks
+                    const input = state.input;
+                    const model = state.model ? state.model : "openai/gpt-5"
+                    const instructions = `You are a prompt refiner. Any prompts you receive, you will refine to improve LLM performance. Break down the prompt by Intent, Mood, and Instructions. Do NOT reply or answer the user's message! ONLY refine the prompt.`;
+                    return {
+                        inputText: input,
+                        model: model,
+                        instructions: instructions
+                    }
+                })
+            .addTask(new TABasicTextInference("extract_search_terms", client),
+                (state, outputs) => {
+                    const input = `Original request: ${state.input}\n\nRefined prompt: ${state.refinedInput}`;
+                    const model = state.model ? state.model : "openai/gpt-5"
+                    const instructions = `You will receive the original user request AND an LLM refined version of the prompt. Please use both to extract one short web search query that will retrieve useful results.`;
+                    return {
+                        inputText: input,
+                        model: model,
+                        instructions: instructions
+                    }
+                })
+            .finally((state, outputs) => { // The finally block allows the playlist to react to the last task and to modify state data before the run ends.
+                if (outputs.refine.success) {
+                    state.refinedInput = outputs.refine.data.text
+                } else {
+                    state.refinedInput = "Sorry, an error occurred: " + outputs.refine.error
+                }
+
+                if (outputs.extract_search_terms.success) {
+                    state.searchTerm = outputs.extract_search_terms.data.text
+                }
+            }))
+        .retryLimit(3) // Simple retry rule setters. Also includes .preventRetry() to disable retries entirely and .retryDelayMs(delayMs) to set the delay between retries. Default is infinite retries at 1000ms delay.
+        .addTransition({
+            to: "search_web", // Transitions refer to states by their ident.
+            condition: async (stateData: StateData) => stateData.searchTerm ? true : false,
+            weight: 2 // Weight determines the order in which transitions are tried. Higher weight = higher priority.
+        })
+        .addTransition({
+            to: "generate_response",
+            condition: async (stateData: StateData) => true,
+            weight: 1
+        }),
+        { initial: true } // The machine needs an initial state.
+    )
+    .addState(StateNode.create<StateData>()
+        .setIdent("search_web")
+        .setPlaylist(p => p
+            .addTask(new TASearchOnline("search"),
+            (state, outputs) => {
+                return {
+                    query: state.searchTerm! // We are sure that the searchTerm is not undefined because of the transition condition.
+                }
+            })
+            .finally((state, outputs) => {
+                if(outputs.search.success) {
+                    state.searchResults = outputs.search.data
+                }
+            }))
+        .addTransition({
+            to: "generate_response",
+            condition: async (stateData: StateData) => true,
+            weight: 1
+        })
+    )
+    .addState(StateNode.create<StateData>()
+        .setIdent("generate_response")
+        .setPlaylist(p => p
+            .addTask(new TABasicTextInference("generate_response", client),
+            (state, outputs) => {
+                return {
+                    inputText: state.input,
+                    model: state.model ? state.model : "openai/gpt-5",
+                    instructions: "You will receive a user request and a refined prompt. There may also be search results. Based on the information, please write a professional response to the user's request."
+                }
+            })
+            .finally((state, outputs) => {
+                if(outputs.generate_response.success) {
+                    state.finalResponse = outputs.generate_response.data.text
+                }
+                else {
+                    state.finalResponse = "Sorry, an error occurred: " + outputs.generate_response.error
+                }
+            })
+    ))
+    .finalize({ // Finalize your machine to make it ready to run. Verbose machines emit JSON logs. If you don't provide an ident, a uuidv4 will be generated for it.
+        verbose: true,
+        ident: "web-search-agent"
+    })
 
-A trigger must define the type of event data it produces and can implement a `start` method for background polling.
+// ------------- EXECUTION: -------------
 
-```typescript
-import { Trigger, TriggerEvent } from '@fkws-npm/klonk';
+const state: StateData = { // The state object is mutable and is passed to the machine and playlists.
+    input: "How do I update AMD graphic driver?",
+    model: "openai/gpt-4o-mini"
+}
 
-// 1. Define the shape of the data this trigger produces
-type NewUserEventData = { userId: string; signupDate: Date; };
+// The .run() method executes the machine until it reaches a terminal state
+// (leaf, failed, out of retries, looped back to initial state)
+// and returns the final state. The original state object is also mutated.
+const finalState = await webSearchAgent.run(state)
 
-// 2. Create the class, extending the generic Trigger
-export class TRNewUserSignedUp<IdentType extends string> extends Trigger<
-  IdentType,
-  TriggerEvent<IdentType, NewUserEventData>
-> {
-  constructor(ident: IdentType) {
-    super(ident);
-  }
-
-  // 3. The `start` method is called by the workflow to begin polling
-  async start(): Promise<void> {
-    // Example: check for new users every 10 seconds
-    setInterval(async () => {
-      const newUser = await this.checkForNewUserInDb();
-      if (newUser) {
-        // 4. Add new events to the `this.queue` property
-        this.queue.push({
-          triggerIdent: this.ident,
-          data: newUser
-        });
-      }
-    }, 10000);
-  }
-
-  private async checkForNewUserInDb(): Promise<NewUserEventData | null> {
-    // Your database logic here...
-    return null;
-  }
-}
+console.log(finalState.finalResponse) // The final state is returned.
+// Or simply:
+console.log(state.finalResponse) // original state object is also mutated.
 ```
+</details>