Streaming

Set stream: true and the response becomes a text/event-stream. Events arrive as data: lines separated by blank lines.

Chat completions

Each event is a chat.completion.chunk. The stream ends with data: [DONE].

Chat completion streamCopy

data: {"id":"chatcmpl-nebula","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"role":"assistant","content":"An IDOR"}}]}

data: {"id":"chatcmpl-nebula","object":"chat.completion.chunk","choices":[{"index":0,"delta":{"content":" lets a user reach"}}]}

data: {"id":"chatcmpl-nebula","object":"chat.completion.chunk","choices":[{"index":0,"delta":{},"finish_reason":"stop"}]}

data: [DONE]

Tool calls stream as delta.tool_calls fragments. Concatenate arguments per call index before parsing them as JSON.

Responses

Responses streams use named events: response.created, then response.output_item.added, response.output_text.delta and the matching .done events, one function_call item per tool call, and finally response.completed. If the request fails after the stream starts, the last event is response.failed.

Messages (Anthropic-compatible)

The Messages API streams Anthropic events: message_start, content_block_start, content_block_delta, content_block_stop, message_delta, message_stop, and periodic ping events. Anthropic SDKs handle them unchanged.

Keepalives

While the model is thinking or between chunks, the server sends SSE comment lines to keep connections open through proxies:

Keepalive commentCopy

: keepalive

Comment lines start with a colon. Standard SSE clients ignore them; if you parse the stream yourself, skip any line that starts with :.

Errors after the stream starts

The HTTP status is already 200 once streaming begins, so a later failure arrives in-band as a final data: event with an error object, followed by data: [DONE]. See Errors.

Security

Streamed and non-streamed output is scrubbed of BreachLine’s own credentials (bl_live_ and bl_test_ keys, and the platform’s provider keys), which are replaced with [REDACTED_SECRET]. By default, other secrets in model output, such as credentials found on a target, are not masked. See Security and data handling.

Practical advice

• Set a read timeout longer than the keepalive interval. Keepalives arrive about every 12 seconds while a stream is idle.
• Treat a stream that closes without [DONE] (or message_stop, or response.completed) as incomplete and retry with an Idempotency-Key.