# Streaming Steps

Streaming steps let a workflow produce output incrementally. Each yielded chunk is written to the database immediately, so if the step fails mid-stream the already-persisted chunks survive and are available on retry.

This is useful for LLM streaming responses, large file processing, and any scenario where partial progress should not be lost.

## The StreamingStep interface

A streaming step implements `Queuety\Contracts\StreamingStep`:

```php
namespace Queuety\Contracts;

interface StreamingStep {
    public function stream( array $state, array $existing_chunks = [] ): \Generator;
    public function on_complete( array $chunks, array $state ): array;
    public function config(): array;
}
```

| Method | Description |
|---|---|
| `stream( $state, $existing_chunks )` | Generator that yields string chunks. Each yield triggers an immediate DB write. |
| `on_complete( $chunks, $state )` | Called after the stream finishes. Receives all chunks (previous + new) and returns data to merge into workflow state. |
| `config()` | Optional configuration. Supports `max_attempts` and `backoff` keys. |

## How chunks are persisted

Every time your generator yields a value, Queuety calls `ChunkStore::append_chunk()` to write it to the `queuety_chunks` table. There is no batching or buffering. One yield equals one INSERT.

The chunk row stores the `job_id`, `workflow_id`, `step_index`, `chunk_index` (zero-based), and the string `content`.

## Resumability

When a streaming step is retried after a failure, the worker loads all previously persisted chunks and passes them as the `$existing_chunks` parameter. Your implementation decides how to resume:

- **Skip already-processed items** by checking the count of existing chunks
- **Offset into a paginated API** using the chunk count as the page offset
- **Restart from scratch** by ignoring `$existing_chunks` entirely (chunks from the previous attempt are still in the DB and will be prepended)

```php
public function stream( array $state, array $existing_chunks = [] ): \Generator {
    $offset = count( $existing_chunks );

    $pages = $this->client->paginate( $state['query'], offset: $offset );

    foreach ( $pages as $page ) {
        yield json_encode( $page );
    }
}
```

## Heartbeats during streaming

Long-running streams should send heartbeats to prevent the stale job detector from reclaiming the job. Call `Heartbeat::beat()` inside your generator loop:

```php
use Queuety\Heartbeat;

public function stream( array $state, array $existing_chunks = [] ): \Generator {
    foreach ( $this->client->stream( $state['prompt'] ) as $i => $chunk ) {
        Heartbeat::beat( [ 'chunks' => $i + 1 ] );
        yield $chunk;
    }
}
```

See [Heartbeats](/docs/features/heartbeats) for details on the Heartbeat API.

## Chunk cleanup

When the stream completes successfully and `on_complete()` returns, the worker calls `ChunkStore::clear_chunks()` to delete all chunks for the job. The merged result from `on_complete()` is the durable record. Chunks are temporary by design.

If the step fails and is retried, chunks are not deleted. They are passed back via `$existing_chunks` on the next attempt.

## Full example: LLM streaming

```php
use Queuety\Contracts\StreamingStep;
use Queuety\Heartbeat;

class StreamLLMResponse implements StreamingStep {

    public function stream( array $state, array $existing_chunks = [] ): \Generator {
        // If we have existing chunks from a previous attempt, we can skip
        // re-requesting. The on_complete method will concatenate everything.
        if ( ! empty( $existing_chunks ) ) {
            return;
        }

        $response = OpenAI::chat()->createStreamed( [
            'model'    => 'gpt-4',
            'messages' => $state['messages'],
        ] );

        foreach ( $response as $i => $chunk ) {
            $text = $chunk->choices[0]->delta->content ?? '';
            if ( '' !== $text ) {
                Heartbeat::beat( [ 'tokens' => $i ] );
                yield $text;
            }
        }
    }

    public function on_complete( array $chunks, array $state ): array {
        return [
            'llm_response' => implode( '', $chunks ),
        ];
    }

    public function config(): array {
        return [ 'max_attempts' => 5 ];
    }
}
```

Use it in a workflow like any other step:

```php
use Queuety\Queuety;

Queuety::workflow( 'ai_chat' )
    ->then( BuildPromptHandler::class )
    ->then( StreamLLMResponse::class )
    ->then( SaveResponseHandler::class )
    ->dispatch( [ 'user_input' => $input ] );
```

## Full example: file processing

```php
use Queuety\Contracts\StreamingStep;
use Queuety\Heartbeat;

class ProcessCSVRows implements StreamingStep {

    public function stream( array $state, array $existing_chunks = [] ): \Generator {
        $file   = fopen( $state['csv_path'], 'r' );
        $offset = count( $existing_chunks );
        $row    = 0;

        // Skip the header.
        fgetcsv( $file );

        while ( $line = fgetcsv( $file ) ) {
            $row++;

            // Skip rows already processed in a previous attempt.
            if ( $row <= $offset ) {
                continue;
            }

            $result = $this->process_row( $line );
            Heartbeat::beat( [ 'row' => $row ] );
            yield json_encode( $result );
        }

        fclose( $file );
    }

    public function on_complete( array $chunks, array $state ): array {
        $results = array_map( 'json_decode', $chunks );
        return [
            'processed_count' => count( $results ),
            'results'         => $results,
        ];
    }

    public function config(): array {
        return [];
    }

    private function process_row( array $line ): array {
        // Your row processing logic here.
        return [ 'id' => $line[0], 'status' => 'ok' ];
    }
}
```

## ChunkStore API

The `ChunkStore` class manages chunk persistence. Access it via `Queuety::chunk_store()`.

```php
$store = Queuety::chunk_store();
```

| Method | Returns | Description |
|---|---|---|
| `get_chunks( int $job_id )` | `string[]` | Fetch all chunks for a job, ordered by chunk index |
| `append_chunk( int $job_id, int $chunk_index, string $content, ?int $workflow_id, ?int $step_index )` | `void` | Append a single chunk |
| `clear_chunks( int $job_id )` | `void` | Delete all chunks for a job |
| `chunk_count( int $job_id )` | `int` | Count the number of chunks for a job |
| `get_accumulated( int $job_id )` | `string` | Concatenate all chunks into one string |

You normally do not need to call `ChunkStore` directly. The worker handles persistence and cleanup automatically when processing a `StreamingStep`. The API is available for advanced use cases such as inspecting in-progress chunks or manual cleanup.