Streaming and SSE | SRouter Docs

Client streaming

OpenAI-compatible chat requests can set stream: true. The gateway returns Server-Sent Events and forwards provider output as it becomes available.

curl -N http://localhost:3000/v1/chat/completions \
  -H "Authorization: Bearer sr-live-your_virtual_key" \
  -H "Content-Type: application/json" \
  -d '{"model":"antigravity/gemini-3.7-flash-high","messages":[{"role":"user","content":"Stream this"}],"stream":true}'

Runtime ownership

packages/executors/src/sse.ts parses upstream SSE frames.
packages/executors/src/stream-utils.ts normalizes stream handling.
Provider executors adapt upstream event shapes.
packages/translator maps usage and protocol fields.
apps/api/src/services/usageEvents.ts publishes usage/log events.

Keep streaming code at these boundaries. Do not add provider-specific parsing to an API route.

Dashboard events

The dashboard uses the logs event stream for live request visibility. It is distinct from a model completion stream: one serves the client response, the other serves operator telemetry. The web hook is apps/web/src/hooks/useLogsStream.ts.

Failure behavior

Provider retries and fallback decisions happen before or around execution according to the provider’s contract. A stream can terminate with an upstream error after headers have been sent; clients should treat an incomplete stream as a failed request and inspect the gateway logs for the provider and request context.