> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ai-stats.phaseo.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Streaming

> Stream model output in real time using SSE on core text endpoints.

Streaming lets your UI render tokens as they are generated instead of waiting for a full response.

## Supported endpoints

* `POST /v1/responses`
* `POST /v1/chat/completions`
* `POST /v1/messages`

## Enable streaming

Set `stream: true` in the request body.

```bash theme={null}
curl https://api.phaseo.app/v1/responses \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-5-nano",
    "input": "Write a short greeting.",
    "stream": true
  }'
```

## SSE frame shape

Streams are returned as SSE frames:

```text theme={null}
data: {"id":"resp_...","status":"in_progress",...}
data: {"type":"response.output_text.delta","delta":"Hello"}
data: {"type":"response.completed",...}
data: [DONE]
```

## Error handling during streams

* If a request fails before streaming starts, you receive a normal JSON error response.
* If a request fails after partial output, treat the stream as incomplete and surface a retry path in your UI.
* Always log `generation_id` (when present) for support and correlation.

## Cancellation

Use cancellation controls (`AbortController` in JS, request timeout in backend workers) so abandoned streams do not consume unnecessary capacity.

## Known limitation

At the current gateway request-validation layer, `stream: true` with tool-calling is rejected. Use non-streaming for tool-calling loops.

## Related pages

* [Responses endpoint](./endpoint/responses.mdx)
* [Chat Completions endpoint](./endpoint/chat-completions.mdx)
* [Messages endpoint](./endpoint/anthropic-messages.mdx)
* [Errors and Debugging](./errors.mdx)

If you are implementing streaming support as an agent:

* Use repository skills for SSE parsing, retries, and cancellation handling.
* Prefer endpoint-specific schemas before adding provider-specific assumptions.
* Keep incremental output rendering separate from final-state persistence.