Skip to main content
POST /chat streams the agent’s response as Server-Sent Events (AG-UI protocol frames). The adapter is responsible only for the wire mechanics — headers first, frames verbatim, backpressure honored, teardown on disconnect. The event semantics live in the AG-UI chat reference.

​
What the adapter writes

On a chat request the adapter:
  1. Flushes headers immediately (X-Accel-Buffering: no, Cache-Control: no-cache, the AG-UI content type) so the browser starts receiving events before the first token.
  2. Writes each frame verbatim — data: ...\n\n payload lines from the runtime, plus : heartbeat comment lines on the runtime’s heartbeat cadence (default 15s) to keep intermediaries from timing the connection out.
  3. Honors backpressure — if res.write() returns false, it waits for the socket’s drain event before writing the next frame, so a slow client can’t balloon memory.
  4. Unsubscribes on disconnect — if the client hangs up mid-stream, the adapter calls iterator.return() on the runtime subscription immediately. No zombie pumps, no leaked timers.
Your client sees a normal EventSource / fetch-stream:

​
Reconnect & resume

The POST /chat response carries an x-persona-run-id header. If the connection drops, the client can reattach to the same run and keep receiving the tail of the stream:
The runtime keeps finished runs resumable for runGraceMs (default 5 minutes) in-process. Because resume state lives in the runtime instance’s memory, route all chat traffic for a given user to the same instance (sticky sessions) until multi-instance resume ships — see the Reconnect & resume reference for the full mechanics and the honest limitation.

​
Binary streams

File downloads (GET /files/:id) and any other kind: 'binary' response stream through the same writer, chunk by chunk, without buffering the whole file.

​
Disconnect teardown, in detail

The adapter registers a close handler on the response that calls return() on the runtime’s subscription iterator. Because the runtime’s subscriptions return a fresh iterator per [Symbol.asyncIterator]() call, the adapter pins the loop to the single iterator it holds — so return() always cancels the exact subscription being consumed. A mid-stream disconnect therefore stops the upstream run within one event loop tick. After a normal completion (or after the close teardown), the close listener is removed — there are no lingering listeners across requests.

​
Example: live token streaming

Each AG-UI event arrives as data: {...} — the messages/partial events carry token deltas; the terminal messages/complete event ends the run. See the AG-UI event reference for the full event taxonomy.