# Conversations

Conversations handle incoming messages and route them based on channel. They're the primary interface between users and your bot, providing AI-powered responses and integration with workflows.

## Basic Concepts

### What are Conversations?

- **Channel-specific**: Each conversation handles a specific channel (chat, Slack, etc.)
- **Stateful**: Maintain state per conversation
- **AI-powered**: Built-in support for AI execution with tools and knowledge
- **Interactive**: Can trigger workflows and handle workflow requests

### File Location

- **Location**: `src/conversations/*.ts`
- **Auto-registration**: Files automatically register for their specified channel

## Creating Conversations

### Basic Structure

```typescript
import { Conversation, z } from '@botpress/runtime'

export const Chat = new Conversation({
  channel: 'chat.channel', // Required: channel ID

  // Optional: Per-conversation state
  state: z.object({
    count: z.number().default(0),
    activeWorkflowId: z.string().optional(),
  }),

  async handler({ message, state, conversation, execute, type, event, request, completion }) {
    // Access conversation state via the state parameter
    state.count += 1

    // Send messages via conversation instance
    await conversation.send({
      type: 'text',
      payload: { text: 'Hello!' },
    })

    // AI-powered response
    await execute({
      instructions: 'You are a helpful assistant',
      tools: [searchTool],
      knowledge: [MyKnowledgeBase],
    })
  },
})
```

### Handler Parameters

```typescript
async handler({ message, state, conversation, execute, type, event, request, completion, client }) {
  // message - Incoming message object
  // state - Conversation state (mutable, automatically tracked)
  // conversation - Conversation instance with send() and other methods
  // execute - AI execution function
  // type - Event type: "message" | "event" | "workflow_request" | "workflow_callback"
  // event - Raw event (typed per handler type — see workflow sections below)
  // request - Workflow request object (when type === "workflow_request")
  // completion - Workflow callback object (when type === "workflow_callback")
  // client - BotClient for API calls
}
```

> **Important:** Inside conversation handlers, always use the `conversation` parameter provided by the handler. Do **not** use `context.get("conversation")` — that pattern is for Actions, Tools, and Triggers that may optionally run within a conversation context. The handler parameter is already properly scoped and typed.

## Common Channel IDs

Different platforms have different channel IDs:

```typescript
// Botpress Chat
channel: 'chat.channel'

// Slack
channel: 'slack.dm' // Direct messages
channel: 'slack.channel' // Channel messages

// Discord
channel: 'discord.dm'
channel: 'discord.channel'

// WhatsApp
channel: 'whatsapp.channel'

// Webchat
channel: 'webchat.channel'

// Teams
channel: 'teams.dm'
channel: 'teams.channel'
```

## Message Handling

### Basic Message Processing

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ message, conversation, execute }) {
    // Check message type
    if (message?.type === 'text') {
      const text = message.payload.text

      // Command handling
      if (text.startsWith('/help')) {
        await conversation.send({
          type: 'text',
          payload: { text: 'Available commands:\n/help - Show help\n/status - Check status' },
        })
        return
      }

      // Regular message - use AI
      await execute({
        instructions: 'Help the user with their request',
      })
    }

    // Handle other message types
    if (message?.type === 'image') {
      await conversation.send({
        type: 'text',
        payload: { text: 'I received your image!' },
      })
    }
  },
})
```

See **[Messages](./messages.md)** for complete guide on all message types, metadata, and sending patterns.

> **Route natural language to `execute()` — don't hand-roll parsers.** Match exact strings only for explicit slash-commands (`/help`) or structured event `type`s. For anything users phrase in their own words, call `execute({ instructions })` and pull structured fields with `adk.zai.extract` (see [zai-agent-reference.md](./zai-agent-reference.md)); keyword/regex/`startsWith` matching silently misses real phrasings.

## Custom & Proactive Events

To push a message into a conversation from an external trigger (e.g. an announcement or status update), handle a custom event **inside the Conversation** — declare it in `events` and branch on `type === 'event'`. Don't build a separate `Trigger` + `client.createMessage`; an integration can't post messages as itself.

```typescript
export const Chat = new Conversation({
  channel: '*',
  events: ['chat:custom'],
  async handler({ type, event, conversation }) {
    if (type === 'event') {
      // The pushed data is nested at event.payload.payload. On channel '*' the handler
      // types event as `unknown`, so read it with a cast (or use a concrete channel).
      const { orderId } = (event as { payload: { payload: { orderId: string } } }).payload.payload
      await conversation.send({ type: 'text', payload: { text: `Order ${orderId} shipped!` } })
      return
    }
    // ...handle messages as normal
  },
})
```

To test a pushed event without an external trigger, fire it from an eval `event` turn (`event: { payload: {...} }`) and run `adk evals` — `adk chat --single` only sends user text. See the `adk-evals` skill.

## Conversation Instance Methods

The `conversation` object provides methods for interacting with the current conversation.

### conversation.send()

Send messages to the conversation:

```typescript
// Send text message
await conversation.send({
  type: 'text',
  payload: { text: 'Hello!' },
})

// Send choice message
await conversation.send({
  type: 'choice',
  payload: {
    text: 'Choose an option:',
    options: [
      { label: 'Option 1', value: 'opt1' },
      { label: 'Option 2', value: 'opt2' },
    ],
  },
})
```

See **[Messages](./messages.md)** for all message types and payloads.

### conversation.startTyping()

Show typing indicator to the user:

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ conversation }) {
    // Start typing indicator
    await conversation.startTyping()

    // Do some work (API call, processing, etc.)
    await fetchData()

    // Send message (stops typing automatically)
    await conversation.send({
      type: 'text',
      payload: { text: "Here's the data!" },
    })
  },
})
```

### conversation.stopTyping()

Manually stop typing indicator:

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ conversation }) {
    await conversation.startTyping()

    try {
      const result = await riskyOperation()
      await conversation.send({
        type: 'text',
        payload: { text: result },
      })
    } catch (error) {
      // Stop typing if sending fails
      await conversation.stopTyping()
      throw error
    }
  },
})
```

### conversation.tags

Access and modify conversation tags:

```typescript
// Read tags
const priority = conversation.tags.priority
const category = conversation.tags.category

// Set tags
conversation.tags.priority = 'high'
conversation.tags.category = 'support'
```

See **[Tags](./tags.md)** for complete tag documentation.

### Conversation Properties

```typescript
// Conversation ID
const id = conversation.id

// Channel information
const channel = conversation.channel // e.g., "webchat.channel"
const integration = conversation.integration // e.g., "webchat"
const alias = conversation.alias // Integration alias
```

## AI Execution

### Basic AI Integration

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ message, execute }) {
    await execute({
      instructions: 'You are a helpful customer support assistant',

      // Optional: Add tools
      tools: [searchTool, createTicketTool],

      // Optional: Add knowledge bases
      knowledge: [DocsKnowledgeBase, FAQKnowledgeBase],

      // Optional: Model configuration
      model: 'openai:gpt-4o',
      temperature: 0.7,

      // Optional: Execution hooks
      hooks: {
        onTrace: ({ trace }) => console.log(trace),
        onIterationEnd: async (iteration) => {
          if (iteration.isFailed()) {
            console.log('Iteration failed:', iteration.error)
          }
        },
      },
    })
  },
})
```

See **[Actions](./actions.md#converting-actions-to-tools)** for converting actions to tools with `.asTool()`.

## State Management

### Conversation State

Each conversation maintains its own state:

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',

  state: z.object({
    messageCount: z.number().default(0),
    userName: z.string().optional(),
    preferences: z
      .object({
        language: z.string().default('en'),
        theme: z.string().default('light'),
      })
      .default({}),
  }),

  async handler({ message, state, conversation }) {
    // Access and modify state via the state parameter
    state.messageCount += 1

    if (!state.userName && message?.type === 'text') {
      state.userName = message.payload.text
      await conversation.send({
        type: 'text',
        payload: { text: `Nice to meet you, ${state.userName}!` },
      })
    }

    // State persists across messages
    if (state.messageCount > 10) {
      await conversation.send({
        type: 'text',
        payload: { text: 'Thanks for being an active user!' },
      })
    }
  },
})
```

## Workflow Integration

### Starting Workflows from Conversations

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',
  state: z.object({
    activeWorkflowId: z.string().optional(),
  }),

  async handler({ message, state, conversation }) {
    const text = message?.payload.text || ''

    if (text.startsWith('/process')) {
      // Start workflow with conversation context
      const instance = await ProcessWorkflow.start({
        conversationId: conversation.id, // Critical for communication!
        userId: user.id,
        data: text.substring(8),
      })

      // Store workflow ID
      state.activeWorkflowId = instance.id

      await conversation.send({
        type: 'text',
        payload: { text: `Started processing: ${instance.id}` },
      })
      return
    }

    if (text === '/status' && state.activeWorkflowId) {
      // Check workflow status
      const instance = await ProcessWorkflow.getOrCreate({
        key: state.activeWorkflowId,
        input: {
          /* required input */
        },
      })

      await conversation.send({
        type: 'text',
        payload: {
          text: `Status: ${instance?.status || 'Not found'}`,
        },
      })
      return
    }
  },
})
```

### Handling Workflow Requests

Workflows can request data from conversations using `step.request()`. The conversation receives a `workflow_request` type with the request object.

**Request Object Structure:**

```typescript
type WorkflowRequest = {
  type: `${WorkflowName}:${RequestName}` // e.g., "processOrder:email"
  workflow: BaseWorkflowInstance // Workflow instance
  step: string // Step name
}
```

**Using the Type Discriminant (Recommended)**

When `type === "workflow_request"`, both `request` and `event` are properly typed:

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ type, request, event, conversation, message }) {
    if (type === 'workflow_request') {
      // request is typed as WorkflowRequest
      // event is typed as WorkflowDataRequestEventType
      // event.payload.message contains the prompt from step.request()

      await conversation.send({
        type: 'text',
        payload: { text: event.payload.message },
      })

      // Check specific request type
      if (request.type === 'processOrder:email') {
        const email = message?.payload.text || 'user@example.com'
        await request.workflow.provide('email', { email })
      }
      return
    }
  },
})
```

**Legacy: Using `isWorkflowDataRequest` Type Guard (Deprecated)**

> **Deprecated:** Use the `type === "workflow_request"` discriminant instead. It provides the same typed event access plus the `request` object for additional context.

```typescript
import { isWorkflowDataRequest } from '@botpress/runtime'

export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ event, conversation }) {
    // ⚠️ Deprecated — prefer type === "workflow_request"
    if (isWorkflowDataRequest(event)) {
      await conversation.send({
        type: 'text',
        payload: { text: event.payload.message },
      })

      const userInput = getUserInput()
      await MyWorkflow.provide(event, { data: userInput })
      return
    }
  },
})
```

### Handling Workflow Callbacks (Completion Events)

When a workflow completes, fails, is canceled, or times out, the conversation receives a `workflow_callback` event. Use the `completion` object to inspect the result.

**Completion Object Structure:**

```typescript
type WorkflowCallback = {
  type: string // Workflow name
  workflow: BaseWorkflowInstance // Workflow instance
  status: 'completed' | 'failed' | 'canceled' | 'timed_out'
  output?: z.infer<WorkflowOutput> // Workflow output (if completed)
  error?: string // Error message (if failed)
}
```

**Using the Type Discriminant (Recommended)**

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ type, completion, event, conversation }) {
    if (type === 'workflow_callback') {
      // completion is typed as WorkflowCallback
      // event is typed as WorkflowCallbackEventType

      switch (completion.status) {
        case 'completed':
          await conversation.send({
            type: 'text',
            payload: {
              text: `Workflow "${completion.type}" completed successfully!`,
            },
          })
          // Access workflow output
          if (completion.output) {
            console.log('Output:', completion.output)
          }
          break

        case 'failed':
          await conversation.send({
            type: 'text',
            payload: {
              text: `Workflow "${completion.type}" failed: ${completion.error}`,
            },
          })
          break

        case 'canceled':
          await conversation.send({
            type: 'text',
            payload: { text: `Workflow "${completion.type}" was canceled.` },
          })
          break

        case 'timed_out':
          await conversation.send({
            type: 'text',
            payload: { text: `Workflow "${completion.type}" timed out.` },
          })
          break
      }
      return
    }
  },
})
```

**Legacy: Using `isWorkflowCallback` Type Guard (Deprecated)**

> **Deprecated:** Use the `type === "workflow_callback"` discriminant instead. It provides a typed `completion` object with workflow instance, status, output, and error information.

```typescript
import { isWorkflowCallback } from '@botpress/runtime'

export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ event, conversation }) {
    // ⚠️ Deprecated — prefer type === "workflow_callback"
    if (isWorkflowCallback(event)) {
      console.log('Workflow completed:', event.payload)
    }
  },
})
```

### Combined Workflow Event Handling

A single conversation handler can handle all workflow event types:

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',
  state: z.object({
    activeWorkflowId: z.string().optional(),
  }),

  async handler({ type, message, request, completion, event, conversation, execute }) {
    // Handle workflow data requests
    if (type === 'workflow_request') {
      await conversation.send({
        type: 'text',
        payload: { text: event.payload.message },
      })
      // Wait for user input, then provide it back
      return
    }

    // Handle workflow completion/failure
    if (type === 'workflow_callback') {
      if (completion.status === 'completed') {
        await conversation.send({
          type: 'text',
          payload: { text: `Done! Result: ${JSON.stringify(completion.output)}` },
        })
      } else {
        await conversation.send({
          type: 'text',
          payload: { text: `Workflow ended with status: ${completion.status}` },
        })
      }
      return
    }

    // Handle regular messages
    if (type === 'message') {
      await execute({
        instructions: 'You are a helpful assistant',
      })
    }
  },
})
```

See **[Workflows](./workflows.md#conversation-communication)** for complete workflow request patterns including state management and multiple request types.

## Command Handling

### Implementing Commands

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ message, conversation, execute }) {
    if (message?.type !== 'text') return

    const text = message.payload.text
    const command = text.split(' ')[0].toLowerCase()
    const args = text.substring(command.length).trim()

    switch (command) {
      case '/help':
        await conversation.send({
          type: 'text',
          payload: {
            text: `Available commands:
/help - Show this help
/search <query> - Search knowledge base
/refresh - Refresh knowledge base
/workflow <data> - Start workflow`,
          },
        })
        break

      case '/search':
        const results = await MyKnowledgeBase.search({ query: args, limit: 5 })
        await conversation.send({
          type: 'text',
          payload: {
            text: results.length > 0 ? `Found ${results.length} results:\n${results.join('\n')}` : 'No results found',
          },
        })
        break

      case '/refresh':
        await MyKnowledgeBase.refresh()
        await conversation.send({
          type: 'text',
          payload: { text: 'Knowledge base refreshed!' },
        })
        break

      case '/workflow':
        const instance = await MyWorkflow.getOrCreate({
          key: user.id,
          input: { userId: user.id, data: args },
        })
        await conversation.send({
          type: 'text',
          payload: { text: `Workflow ${instance.id} started` },
        })
        break

      default:
        // Not a command, use AI
        await execute({
          instructions: 'You are a helpful assistant',
          knowledge: [MyKnowledgeBase],
        })
    }
  },
})
```

## Advanced Patterns

### Multi-Channel Support

Create separate conversations for different channels:

```typescript
// src/conversations/slack-dm.ts
export const SlackDM = new Conversation({
  channel: 'slack.dm',

  async handler({ message, execute }) {
    // Slack-specific handling
    if (message?.payload.thread_ts) {
      // Handle threaded messages
    }

    await execute({
      instructions: 'You are a Slack assistant. Use Slack markdown formatting.',
    })
  },
})

// src/conversations/slack-channel.ts
export const SlackChannel = new Conversation({
  channel: 'slack.channel',

  async handler({ message, execute }) {
    // Only respond when mentioned
    if (!message?.payload.text?.includes('@bot')) {
      return // Don't respond
    }

    await execute({
      instructions: 'You are in a public Slack channel. Be concise.',
    })
  },
})
```

### Conversation Context

Maintain context across messages:

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',
  state: z.object({
    context: z
      .array(
        z.object({
          role: z.string(),
          content: z.string(),
        })
      )
      .default([]),
  }),

  async handler({ message, state, execute }) {
    // Add user message to context
    if (message?.type === 'text') {
      state.context.push({
        role: 'user',
        content: message.payload.text,
      })
    }

    // Keep last 10 messages
    if (state.context.length > 10) {
      state.context = state.context.slice(-10)
    }

    // Use context in AI execution
    await execute({
      instructions: `You are a helpful assistant.
Previous context:
${state.context.map((m) => `${m.role}: ${m.content}`).join('\n')}`,
    })
  },
})
```

### Context API & Chat Instance

Access the chat instance to read conversation history, register custom components, and build adaptive experiences.

#### Accessing the Chat Instance

```typescript
import { context, Conversation } from '@botpress/runtime'

export const Chat = new Conversation({
  channel: 'webchat.channel',

  async handler({ message, execute, conversation }) {
    // Get the chat instance
    const chat = context.get('chat')

    // Fetch conversation transcript
    const transcript = await chat.fetchTranscript()

    // Analyze conversation history
    const messageCount = transcript.length
    const hasImages = transcript.some((msg) => msg.role === 'user' && msg.attachments && msg.attachments.length > 0)

    console.log(`Conversation has ${messageCount} messages, includes images: ${hasImages}`)
  },
})
```

#### Adaptive Model Selection

Choose models dynamically based on conversation content (e.g., handle images with vision models):

```typescript
import { context, Conversation } from '@botpress/runtime'

export const Chat = new Conversation({
  channel: 'webchat.channel',

  async handler({ message, execute, conversation }) {
    const chat = context.get('chat')
    const transcript = await chat.fetchTranscript()

    // Check if any user messages contain images
    const hasImages = transcript.some((msg) => msg.role === 'user' && msg.attachments?.length > 0)

    // Select appropriate model
    const model = hasImages
      ? 'openai:gpt-4.1' // Vision-capable model for images
      : 'cerebras:gpt-oss-120b' // Fast, cost-effective for text-only

    await execute({
      instructions: 'You are a helpful assistant...',
      model,
      tools: [myTool],
    })
  },
})
```

**Use Cases:**

- **Cost Optimization**: Use cheaper/faster models for text-only conversations
- **Capability Matching**: Switch to vision models only when images are present
- **Context-Aware Responses**: Adapt behavior based on conversation history
- **Quality Control**: Use more powerful models for complex or lengthy conversations

#### Custom UI Components

Custom components are React `.bp.tsx` files that render rich UI in webchat. Register them in `src/components/index.ts` with `CustomComponent`, then list them in a conversation's `components` array:

```typescript
import { Conversation } from '@botpress/runtime'
import { TicketCardComponent } from '../components'

export const Chat = new Conversation({
  channel: 'webchat.channel',

  components: [TicketCardComponent],

  async handler({ execute, conversation }) {
    // The LLM can yield TicketCard autonomously
    await execute({
      instructions: 'Always use the TicketCard component to display ticket details.',
      tools: [lookupTicket],
    })
  },
})
```

You can also send a component explicitly:

```typescript
import { WelcomeBannerComponent } from '../components'

await conversation.send({
  type: 'customComponent',
  payload: {
    component: WelcomeBannerComponent,
    props: {},
  },
})
```

See the component registry reference for the full component creation flow (`.bp.tsx` files, metadata, and registration).

**Component Use Cases:**

- **Rich Answers**: Formatted responses with citations and styling
- **Interactive Elements**: Buttons, forms, and action cards
- **Data Visualization**: Charts, tables, and structured data display
- **Custom Workflows**: Multi-step interactions with state

#### Chat Instance Methods

The chat instance provides methods for managing conversation transcripts and components.

**Transcript Management:**

```typescript
import { context } from '@botpress/runtime'

const chat = context.get('chat')

// Fetch transcript (loads from API)
const transcript = await chat.fetchTranscript()

// Get transcript (returns cached copy)
const cached = await chat.getTranscript()

// Set/replace entire transcript
await chat.setTranscript([
  { id: '1', role: 'user', content: 'Hello' },
  { id: '2', role: 'assistant', content: 'Hi there!' },
])

// Clear all messages
await chat.clearTranscript()

// Prepend messages to beginning
await chat.prependToTranscript([{ id: '0', role: 'summary', content: 'Previous conversation summary...' }])

// Remove specific message by ID
const removed = await chat.removeMessage('msg-123')

// Remove messages by predicate
const count = await chat.removeMessages((item) => item.role === 'assistant' && item.content.includes('error'))

// Compact transcript (summarize/compress)
await chat.compactTranscript()

// Save transcript changes
await chat.saveTranscript()
```

**Component Management:**

```typescript
const chat = context.get('chat')

// Register component with handler
chat.registerComponent({
  component: myComponent,
  handler: async (props) => {
    // Handle component rendering
  },
})

// Remove component by name
chat.removeComponent('Answer')

// Get all registered components
const components = await chat.getComponents()
```

**Transcript Item Types:**

```typescript
type TranscriptUserMessage = {
  id: string
  role: 'user'
  content: string
  name?: string
  createdAt?: string
  attachments?: Array<{ type: 'image'; url: string }>
}

type TranscriptAssistantMessage = {
  id: string
  role: 'assistant'
  content: string
  name?: string
  createdAt?: string
}

type TranscriptEventMessage = {
  id: string
  role: 'event'
  name: string
  payload: unknown
  createdAt?: string
}

type TranscriptSummaryMessage = {
  id: string
  role: 'summary'
  content: string
}
```

#### Transcript Analysis Patterns

```typescript
import { context, Conversation } from '@botpress/runtime'

export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ message, execute }) {
    const chat = context.get('chat')
    const transcript = await chat.fetchTranscript()

    // Detect conversation patterns
    const userMessages = transcript.filter((msg) => msg.role === 'user')
    const hasRepeatedQuestion = userMessages.some(
      (msg, idx) => idx > 0 && msg.content === userMessages[idx - 1].content
    )

    if (hasRepeatedQuestion) {
      // User asked the same question twice - may need clarification
      await execute({
        instructions: `The user seems frustrated or confused.
        Provide extra clarification and ask if they need more help.`,
        tools: [searchTool],
      })
    } else {
      // Normal conversation flow
      await execute({
        instructions: 'You are a helpful assistant',
        tools: [searchTool],
      })
    }
  },
})
```

**Analysis Patterns:**

- **Detect frustration**: Repeated questions, negative sentiment
- **Identify complexity**: Long conversations may need escalation
- **Track engagement**: Message frequency and length analysis
- **Content-based routing**: Images, files, code snippets

### Error Handling

```typescript
export const Chat = new Conversation({
  channel: 'chat.channel',

  async handler({ message, conversation, execute }) {
    try {
      await execute({
        instructions: 'You are a helpful assistant',
        tools: [riskyTool],
      })
    } catch (error) {
      console.error('Execution failed:', error)

      // Send error message to user
      await conversation.send({
        type: 'text',
        payload: {
          text: 'I apologize, but I encountered an error. Please try again later.',
        },
      })

      // Optionally notify admins
      await notifyAdmins({
        error: error.message,
        conversationId: conversation.id,
      })
    }
  },
})
```

## Best Practices

### 1. Handle All Message Types

```typescript
async handler({ message, conversation }) {
  switch (message?.type) {
    case "text":
      // Handle text
      break;
    case "image":
      // Handle image
      break;
    case "choice":
      // Handle choice selection
      break;
    default:
      // Handle unknown types
  }
}
```

### 2. Validate Input

```typescript
if (message?.type === 'text') {
  const text = message.payload.text?.trim()
  if (!text || text.length > 1000) {
    await conversation.send({
      type: 'text',
      payload: { text: 'Please provide a valid message (1-1000 characters)' },
    })
    return
  }
}
```

### 3. Use Early Returns

```typescript
async handler({ message, state, conversation, execute }) {
  // Handle commands first
  if (message?.payload.text?.startsWith("/")) {
    // Handle command inline
    return; // Early return
  }

  // Handle special cases
  if (state.waitingForInput) {
    // Handle input
    return; // Early return
  }

  // Default AI handling
  await execute({ instructions: "..." });
}
```

### 4. Provide Clear Instructions

```typescript
await execute({
  instructions: `You are a customer support assistant for ACME Corp.

Guidelines:
- Be helpful and professional
- Keep responses concise (under 100 words)
- Always verify customer identity before sharing sensitive info
- Escalate to human agent if customer seems frustrated

Current user: ${user.name}
Account type: ${user.accountType}`,
})
```

### 5. Clean Up State

```typescript
// Reset state when conversation ends
async handler({ message, state, conversation }) {
  if (message?.payload.text === "/end") {
    // Reset state properties
    state.count = 0;
    state.userName = undefined;
    state.activeWorkflowId = undefined;

    await conversation.send({
      type: "text",
      payload: { text: "Conversation ended. State cleared." }
    });
  }
}
```