fix(streaming): terminate SSE with [DONE] and keep-alives

src/server/mod.rs

@@ -6,6 +6,7 @@ use axum::{
     routing::{get, post},
 };
 use futures::stream::StreamExt;
+use std::time::Duration;
 use sqlx;
 use std::sync::Arc;
 use tracing::{info, warn};
@@ -279,7 +280,21 @@ async fn chat_completions(
                     }
                 });
 
-                Ok(Sse::new(sse_stream).into_response())
+                // Many OpenAI-compatible clients expect a terminal [DONE] marker.
+                // Emit it when the upstream stream ends to avoid clients treating
+                // the response as incomplete.
+                let done = futures::stream::once(async { Ok(Event::default().data("[DONE]")) });
+                let out = sse_stream.chain(done);
+
+                Ok(
+                    Sse::new(out)
+                        .keep_alive(
+                            axum::response::sse::KeepAlive::new()
+                                .interval(Duration::from_secs(15))
+                                .text(": keep-alive"),
+                        )
+                        .into_response(),
+                )
             }
             Err(e) => {
                 // Record provider failure