v4.5.5

changelog.md

@@ -1,5 +1,68 @@
 # Changelog
 
+## 2026-03-16 - 4.5.5 - fix(remoteingress-core)
+wait for hub-to-client draining before cleanup and reliably send close frames
+
+- switch CLOSE frame delivery on the data channel from try_send to send().await to avoid dropping it when the channel is full
+- delay stream cleanup until the hub-to-client task finishes or times out so large downstream responses continue after upload EOF
+- add a bounded 5-minute wait for download draining to prevent premature termination of asymmetric transfers such as git fetch
+
+## 2026-03-15 - 4.5.4 - fix(remoteingress-core)
+preserve stream close ordering and add flow-control stall timeouts
+
+- Send CLOSE and CLOSE_BACK frames on the data channel so they arrive after the final stream data frames.
+- Log and abort stalled upload and download paths when flow-control windows stay empty for 120 seconds.
+- Apply a 60-second timeout when writing buffered stream data to the upstream connection to prevent hung streams.
+
+## 2026-03-15 - 4.5.3 - fix(remoteingress-core)
+prioritize control frames over data in edge and hub tunnel writers
+
+- Split tunnel/frame writers into separate control and data channels in edge and hub
+- Use biased select loops so PING, PONG, WINDOW_UPDATE, OPEN, and CLOSE frames are sent before data frames
+- Route stream data through dedicated data channels while keeping OPEN, CLOSE, and flow-control updates on control channels to prevent keepalive starvation under load
+
+## 2026-03-15 - 4.5.2 - fix(remoteingress-core)
+improve stream flow control retries and increase channel buffer capacity
+
+- increase per-stream mpsc channel capacity from 128 to 256 on both edge and hub paths
+- only reset accumulated window update bytes after a successful try_send to avoid dropping flow-control credits when the update channel is busy
+
+## 2026-03-15 - 4.5.1 - fix(protocol)
+increase per-stream flow control window and channel buffers to improve high-RTT throughput
+
+- raise the initial stream window from 256 KB to 4 MB to allow more in-flight data per stream
+- increase edge and hub mpsc channel capacities from 16 to 128 to better absorb throughput under flow control
+
+## 2026-03-15 - 4.5.0 - feat(remoteingress-core)
+add per-stream flow control for edge and hub tunnel data transfer
+
+- introduce WINDOW_UPDATE frame types and protocol helpers for per-stream flow control
+- track per-stream send windows on both edge and hub to limit reads based on available capacity
+- send window updates after downstream writes to reduce channel pressure during large transfers
+
+## 2026-03-15 - 4.4.1 - fix(remoteingress-core)
+prevent stream data loss by applying backpressure and closing saturated channels
+
+- replace non-blocking frame writes with awaited sends in per-stream tasks so large transfers respect backpressure instead of dropping data
+- close and remove streams when back-channel or data channels fill up to avoid TCP stream corruption from silently dropped frames
+
+## 2026-03-03 - 4.4.0 - feat(remoteingress)
+add heartbeat PING/PONG and liveness timeouts; implement fast-reconnect/backoff reset and JS crash-recovery auto-restart
+
+- protocol: add FRAME_PING and FRAME_PONG and unit tests for ping/pong frames
+- edge (Rust): reset backoff after successful connection, respond to PING with PONG, track liveness via deadline and reconnect on timeout, use Duration/Instant helpers
+- hub (Rust): send periodic PING to edges, handle PONGs, enforce liveness timeout and disconnect inactive edges, use tokio interval and time utilities
+- ts: RemoteIngressEdge and RemoteIngressHub: add crash-recovery auto-restart with exponential backoff and max attempts, save/restore config and allowed edges, register/remove exit handlers, ensure stop() marks stopping and cleans up listeners
+- minor API/typing: introduce TAllowedEdge alias and persist allowed edges for restart recovery
+
+## 2026-02-26 - 4.3.0 - feat(hub)
+add optional TLS certificate/key support to hub start config and bridge
+
+- TypeScript: add tls.certPem and tls.keyPem to IHubConfig and include tlsCertPem/tlsKeyPem in startHub bridge command when both are provided
+- TypeScript: extend startHub params with tlsCertPem and tlsKeyPem and conditionally send them
+- Rust: change HubConfig serde attributes for tls_cert_pem and tls_key_pem from skip to default so absent PEM fields deserialize as None
+- Enables optional provisioning of TLS certificate and key to the hub when provided from the JS side
+
 ## 2026-02-26 - 4.2.0 - feat(core)
 expose edge peer address in hub events and migrate writers to channel-based, non-blocking framing with stream limits and timeouts
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@serve.zone/remoteingress",
-  "version": "4.2.0",
+  "version": "4.5.5",
   "private": false,
   "description": "Edge ingress tunnel for DcRouter - accepts incoming TCP connections at network edge and tunnels them to DcRouter SmartProxy preserving client IP via PROXY protocol v1.",
   "main": "dist_ts/index.js",

rust/crates/remoteingress-core/src/edge.rs

@@ -1,16 +1,29 @@
 use std::collections::HashMap;
 use std::sync::atomic::{AtomicU32, Ordering};
 use std::sync::Arc;
+use std::time::Duration;
 use tokio::io::{AsyncBufReadExt, AsyncReadExt, AsyncWriteExt, BufReader};
 use tokio::net::{TcpListener, TcpStream};
-use tokio::sync::{mpsc, Mutex, RwLock};
+use tokio::sync::{mpsc, Mutex, Notify, RwLock};
 use tokio::task::JoinHandle;
+use tokio::time::{Instant, sleep_until};
 use tokio_rustls::TlsConnector;
 use tokio_util::sync::CancellationToken;
 use serde::{Deserialize, Serialize};
 
 use remoteingress_protocol::*;
 
+/// Per-stream state tracked in the edge's client_writers map.
+struct EdgeStreamState {
+    /// Channel to deliver FRAME_DATA_BACK payloads to the hub_to_client task.
+    back_tx: mpsc::Sender<Vec<u8>>,
+    /// Send window for FRAME_DATA (upload direction).
+    /// Decremented by the client reader, incremented by FRAME_WINDOW_UPDATE_BACK from hub.
+    send_window: Arc<AtomicU32>,
+    /// Notifier to wake the client reader when the window opens.
+    window_notify: Arc<Notify>,
+}
+
 /// Edge configuration (hub-host + credentials only; ports come from hub).
 #[derive(Debug, Clone, Deserialize, Serialize)]
 #[serde(rename_all = "camelCase")]
@@ -202,6 +215,13 @@ async fn edge_main_loop(
         // Cancel connection token to kill all orphaned tasks from this cycle
         connection_token.cancel();
 
+        // Reset backoff after a successful connection for fast reconnect
+        let was_connected = *connected.read().await;
+        if was_connected {
+            backoff_ms = 1000;
+            log::info!("Was connected; resetting backoff to {}ms for fast reconnect", backoff_ms);
+        }
+
         *connected.write().await = false;
         let _ = event_tx.try_send(EdgeEvent::TunnelDisconnected);
         active_streams.store(0, Ordering::Relaxed);
@@ -214,7 +234,7 @@ async fn edge_main_loop(
             EdgeLoopResult::Reconnect => {
                 log::info!("Reconnecting in {}ms...", backoff_ms);
                 tokio::select! {
-                    _ = tokio::time::sleep(std::time::Duration::from_millis(backoff_ms)) => {}
+                    _ = tokio::time::sleep(Duration::from_millis(backoff_ms)) => {}
                     _ = cancel_token.cancelled() => break,
                     _ = shutdown_rx.recv() => break,
                 }
@@ -336,23 +356,38 @@ async fn connect_to_hub_and_run(
                 _ = stun_token.cancelled() => break,
             }
             tokio::select! {
-                _ = tokio::time::sleep(std::time::Duration::from_secs(stun_interval)) => {}
+                _ = tokio::time::sleep(Duration::from_secs(stun_interval)) => {}
                 _ = stun_token.cancelled() => break,
             }
         }
     });
 
-    // Client socket map: stream_id -> sender for writing data back to client
-    let client_writers: Arc<Mutex<HashMap<u32, mpsc::Sender<Vec<u8>>>>> =
+    // Client socket map: stream_id -> per-stream state (back channel + flow control)
+    let client_writers: Arc<Mutex<HashMap<u32, EdgeStreamState>>> =
         Arc::new(Mutex::new(HashMap::new()));
 
-    // A5: Channel-based tunnel writer replaces Arc<Mutex<WriteHalf>>
-    let (tunnel_writer_tx, mut tunnel_writer_rx) = mpsc::channel::<Vec<u8>>(4096);
+    // QoS dual-channel tunnel writer: control frames (PONG/WINDOW_UPDATE/CLOSE/OPEN)
+    // have priority over data frames (DATA). Prevents PING starvation under load.
+    let (tunnel_ctrl_tx, mut tunnel_ctrl_rx) = mpsc::channel::<Vec<u8>>(64);
+    let (tunnel_data_tx, mut tunnel_data_rx) = mpsc::channel::<Vec<u8>>(4096);
+    // Legacy alias — control channel for PONG, CLOSE, WINDOW_UPDATE, OPEN
+    let tunnel_writer_tx = tunnel_ctrl_tx.clone();
     let tw_token = connection_token.clone();
     let tunnel_writer_handle = tokio::spawn(async move {
         loop {
             tokio::select! {
-                data = tunnel_writer_rx.recv() => {
+                biased; // control frames always take priority over data
+                ctrl = tunnel_ctrl_rx.recv() => {
+                    match ctrl {
+                        Some(frame_data) => {
+                            if write_half.write_all(&frame_data).await.is_err() {
+                                break;
+                            }
+                        }
+                        None => break,
+                    }
+                }
+                data = tunnel_data_rx.recv() => {
                     match data {
                         Some(frame_data) => {
                             if write_half.write_all(&frame_data).await.is_err() {
@@ -373,6 +408,7 @@ async fn connect_to_hub_and_run(
         &handshake.listen_ports,
         &mut port_listeners,
         &tunnel_writer_tx,
+        &tunnel_data_tx,
         &client_writers,
         active_streams,
         next_stream_id,
@@ -380,6 +416,11 @@ async fn connect_to_hub_and_run(
         connection_token,
     );
 
+    // Heartbeat: liveness timeout detects silent hub failures
+    let liveness_timeout_dur = Duration::from_secs(45);
+    let mut last_activity = Instant::now();
+    let mut liveness_deadline = Box::pin(sleep_until(last_activity + liveness_timeout_dur));
+
     // Read frames from hub
     let mut frame_reader = FrameReader::new(buf_reader);
     let result = loop {
@@ -387,13 +428,34 @@ async fn connect_to_hub_and_run(
             frame_result = frame_reader.next_frame() => {
                 match frame_result {
                     Ok(Some(frame)) => {
+                        // Reset liveness on any received frame
+                        last_activity = Instant::now();
+                        liveness_deadline.as_mut().reset(last_activity + liveness_timeout_dur);
+
                         match frame.frame_type {
                             FRAME_DATA_BACK => {
-                                // A1: Non-blocking send to prevent head-of-line blocking
-                                let writers = client_writers.lock().await;
-                                if let Some(tx) = writers.get(&frame.stream_id) {
-                                    if tx.try_send(frame.payload).is_err() {
-                                        log::warn!("Stream {} back-channel full, dropping frame", frame.stream_id);
+                                // Non-blocking dispatch to per-stream channel.
+                                // With flow control, the sender should rarely exceed the channel capacity.
+                                let mut writers = client_writers.lock().await;
+                                if let Some(state) = writers.get(&frame.stream_id) {
+                                    if state.back_tx.try_send(frame.payload).is_err() {
+                                        log::warn!("Stream {} back-channel full, closing stream", frame.stream_id);
+                                        writers.remove(&frame.stream_id);
+                                    }
+                                }
+                            }
+                            FRAME_WINDOW_UPDATE_BACK => {
+                                // Hub consumed data — increase our send window for this stream (upload direction)
+                                if let Some(increment) = decode_window_update(&frame.payload) {
+                                    if increment > 0 {
+                                        let writers = client_writers.lock().await;
+                                        if let Some(state) = writers.get(&frame.stream_id) {
+                                            let prev = state.send_window.fetch_add(increment, Ordering::Release);
+                                            if prev + increment > MAX_WINDOW_SIZE {
+                                                state.send_window.store(MAX_WINDOW_SIZE, Ordering::Release);
+                                            }
+                                            state.window_notify.notify_one();
+                                        }
                                     }
                                 }
                             }
@@ -412,6 +474,7 @@ async fn connect_to_hub_and_run(
                                         &update.listen_ports,
                                         &mut port_listeners,
                                         &tunnel_writer_tx,
+                                        &tunnel_data_tx,
                                         &client_writers,
                                         active_streams,
                                         next_stream_id,
@@ -420,6 +483,14 @@ async fn connect_to_hub_and_run(
                                     );
                                 }
                             }
+                            FRAME_PING => {
+                                let pong_frame = encode_frame(0, FRAME_PONG, &[]);
+                                if tunnel_writer_tx.try_send(pong_frame).is_err() {
+                                    log::warn!("Failed to send PONG, writer channel full/closed");
+                                    break EdgeLoopResult::Reconnect;
+                                }
+                                log::trace!("Received PING from hub, sent PONG");
+                            }
                             _ => {
                                 log::warn!("Unexpected frame type {} from hub", frame.frame_type);
                             }
@@ -435,6 +506,11 @@ async fn connect_to_hub_and_run(
                     }
                 }
             }
+            _ = &mut liveness_deadline => {
+                log::warn!("Hub liveness timeout (no frames for {}s), reconnecting",
+                    liveness_timeout_dur.as_secs());
+                break EdgeLoopResult::Reconnect;
+            }
             _ = connection_token.cancelled() => {
                 log::info!("Connection cancelled");
                 break EdgeLoopResult::Shutdown;
@@ -460,8 +536,9 @@ async fn connect_to_hub_and_run(
 fn apply_port_config(
     new_ports: &[u16],
     port_listeners: &mut HashMap<u16, JoinHandle<()>>,
-    tunnel_writer_tx: &mpsc::Sender<Vec<u8>>,
-    client_writers: &Arc<Mutex<HashMap<u32, mpsc::Sender<Vec<u8>>>>>,
+    tunnel_ctrl_tx: &mpsc::Sender<Vec<u8>>,
+    tunnel_data_tx: &mpsc::Sender<Vec<u8>>,
+    client_writers: &Arc<Mutex<HashMap<u32, EdgeStreamState>>>,
     active_streams: &Arc<AtomicU32>,
     next_stream_id: &Arc<AtomicU32>,
     edge_id: &str,
@@ -480,7 +557,8 @@ fn apply_port_config(
 
     // Add new ports
     for &port in new_set.difference(&old_set) {
-        let tunnel_writer_tx = tunnel_writer_tx.clone();
+        let tunnel_ctrl_tx = tunnel_ctrl_tx.clone();
+        let tunnel_data_tx = tunnel_data_tx.clone();
         let client_writers = client_writers.clone();
         let active_streams = active_streams.clone();
         let next_stream_id = next_stream_id.clone();
@@ -503,7 +581,8 @@ fn apply_port_config(
                         match accept_result {
                             Ok((client_stream, client_addr)) => {
                                 let stream_id = next_stream_id.fetch_add(1, Ordering::Relaxed);
-                                let tunnel_writer_tx = tunnel_writer_tx.clone();
+                                let tunnel_ctrl_tx = tunnel_ctrl_tx.clone();
+                                let tunnel_data_tx = tunnel_data_tx.clone();
                                 let client_writers = client_writers.clone();
                                 let active_streams = active_streams.clone();
                                 let edge_id = edge_id.clone();
@@ -518,7 +597,8 @@ fn apply_port_config(
                                         stream_id,
                                         port,
                                         &edge_id,
-                                        tunnel_writer_tx,
+                                        tunnel_ctrl_tx,
+                                        tunnel_data_tx,
                                         client_writers,
                                         client_token,
                                     )
@@ -548,8 +628,9 @@ async fn handle_client_connection(
     stream_id: u32,
     dest_port: u16,
     edge_id: &str,
-    tunnel_writer_tx: mpsc::Sender<Vec<u8>>,
-    client_writers: Arc<Mutex<HashMap<u32, mpsc::Sender<Vec<u8>>>>>,
+    tunnel_ctrl_tx: mpsc::Sender<Vec<u8>>,
+    tunnel_data_tx: mpsc::Sender<Vec<u8>>,
+    client_writers: Arc<Mutex<HashMap<u32, EdgeStreamState>>>,
     client_token: CancellationToken,
 ) {
     let client_ip = client_addr.ip().to_string();
@@ -558,33 +639,52 @@ async fn handle_client_connection(
     // Determine edge IP (use 0.0.0.0 as placeholder — hub doesn't use it for routing)
     let edge_ip = "0.0.0.0";
 
-    // Send OPEN frame with PROXY v1 header via writer channel
+    // Send OPEN frame with PROXY v1 header via control channel
     let proxy_header = build_proxy_v1_header(&client_ip, edge_ip, client_port, dest_port);
     let open_frame = encode_frame(stream_id, FRAME_OPEN, proxy_header.as_bytes());
-    if tunnel_writer_tx.send(open_frame).await.is_err() {
+    if tunnel_ctrl_tx.send(open_frame).await.is_err() {
         return;
     }
 
-    // Set up channel for data coming back from hub
+    // Set up channel for data coming back from hub (capacity 16 is sufficient with flow control)
     let (back_tx, mut back_rx) = mpsc::channel::<Vec<u8>>(256);
+    let send_window = Arc::new(AtomicU32::new(INITIAL_STREAM_WINDOW));
+    let window_notify = Arc::new(Notify::new());
     {
         let mut writers = client_writers.lock().await;
-        writers.insert(stream_id, back_tx);
+        writers.insert(stream_id, EdgeStreamState {
+            back_tx,
+            send_window: Arc::clone(&send_window),
+            window_notify: Arc::clone(&window_notify),
+        });
     }
 
     let (mut client_read, mut client_write) = client_stream.into_split();
 
-    // Task: hub -> client
+    // Task: hub -> client (download direction)
+    // After writing to client TCP, send WINDOW_UPDATE to hub so it can send more
     let hub_to_client_token = client_token.clone();
-    let hub_to_client = tokio::spawn(async move {
+    let wu_tx = tunnel_ctrl_tx.clone();
+    let mut hub_to_client = tokio::spawn(async move {
+        let mut consumed_since_update: u32 = 0;
         loop {
             tokio::select! {
                 data = back_rx.recv() => {
                     match data {
                         Some(data) => {
+                            let len = data.len() as u32;
                             if client_write.write_all(&data).await.is_err() {
                                 break;
                             }
+                            // Track consumption for flow control
+                            consumed_since_update += len;
+                            if consumed_since_update >= WINDOW_UPDATE_THRESHOLD {
+                                let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE, consumed_since_update);
+                                if wu_tx.try_send(frame).is_ok() {
+                                    consumed_since_update = 0;
+                                }
+                                // If try_send fails, keep accumulating — retry on next threshold
+                            }
                         }
                         None => break,
                     }
@@ -592,21 +692,45 @@ async fn handle_client_connection(
                 _ = hub_to_client_token.cancelled() => break,
             }
         }
+        // Send final window update for any remaining consumed bytes
+        if consumed_since_update > 0 {
+            let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE, consumed_since_update);
+            let _ = wu_tx.try_send(frame);
+        }
         let _ = client_write.shutdown().await;
     });
 
-    // Task: client -> hub (via writer channel)
+    // Task: client -> hub (upload direction) with per-stream flow control
     let mut buf = vec![0u8; 32768];
     loop {
+        // Wait for send window to have capacity (with stall timeout)
+        loop {
+            let w = send_window.load(Ordering::Acquire);
+            if w > 0 { break; }
+            tokio::select! {
+                _ = window_notify.notified() => continue,
+                _ = client_token.cancelled() => break,
+                _ = tokio::time::sleep(Duration::from_secs(120)) => {
+                    log::warn!("Stream {} upload stalled (window empty for 120s)", stream_id);
+                    break;
+                }
+            }
+        }
+        if client_token.is_cancelled() { break; }
+
+        // Limit read size to available window
+        let w = send_window.load(Ordering::Acquire) as usize;
+        let max_read = w.min(buf.len());
+
         tokio::select! {
-            read_result = client_read.read(&mut buf) => {
+            read_result = client_read.read(&mut buf[..max_read]) => {
                 match read_result {
                     Ok(0) => break,
                     Ok(n) => {
+                        send_window.fetch_sub(n as u32, Ordering::Release);
                         let data_frame = encode_frame(stream_id, FRAME_DATA, &buf[..n]);
-                        // A5: Use try_send to avoid blocking if writer channel is full
-                        if tunnel_writer_tx.try_send(data_frame).is_err() {
-                            log::warn!("Stream {} tunnel writer full, closing", stream_id);
+                        if tunnel_data_tx.send(data_frame).await.is_err() {
+                            log::warn!("Stream {} data channel closed, closing", stream_id);
                             break;
                         }
                     }
@@ -617,18 +741,32 @@ async fn handle_client_connection(
         }
     }
 
-    // Send CLOSE frame (only if not cancelled)
+    // Send CLOSE frame via DATA channel (must arrive AFTER last DATA for this stream).
+    // Use send().await to guarantee delivery (try_send silently drops if channel full).
     if !client_token.is_cancelled() {
         let close_frame = encode_frame(stream_id, FRAME_CLOSE, &[]);
-        let _ = tunnel_writer_tx.try_send(close_frame);
+        let _ = tunnel_data_tx.send(close_frame).await;
     }
 
-    // Cleanup
+    // Wait for the download task (hub → client) to finish draining all buffered
+    // response data. Upload EOF just means the client is done sending; the download
+    // must continue until all response data has been written to the client.
+    // This is critical for asymmetric transfers like git fetch (small request, large response).
+    // The download task will exit when:
+    //   - back_rx returns None (back_tx dropped below after await, or hub sent CLOSE_BACK)
+    //   - client_write fails (client disconnected)
+    //   - client_token is cancelled
+    let _ = tokio::time::timeout(
+        Duration::from_secs(300),  // 5 min max wait for download to finish
+        &mut hub_to_client,
+    ).await;
+
+    // Now safe to clean up — download has finished or timed out
     {
         let mut writers = client_writers.lock().await;
         writers.remove(&stream_id);
     }
-    hub_to_client.abort();
+    hub_to_client.abort(); // No-op if already finished; safety net if timeout fired
     let _ = edge_id; // used for logging context
 }
 

rust/crates/remoteingress-core/src/hub.rs

@@ -1,23 +1,39 @@
 use std::collections::HashMap;
 use std::sync::Arc;
+use std::sync::atomic::{AtomicU32, Ordering};
+use std::time::Duration;
 use tokio::io::{AsyncBufReadExt, AsyncReadExt, AsyncWriteExt, BufReader};
 use tokio::net::{TcpListener, TcpStream};
-use tokio::sync::{mpsc, Mutex, RwLock, Semaphore};
+use tokio::sync::{mpsc, Mutex, Notify, RwLock, Semaphore};
+use tokio::time::{interval, sleep_until, Instant};
 use tokio_rustls::TlsAcceptor;
 use tokio_util::sync::CancellationToken;
 use serde::{Deserialize, Serialize};
 
 use remoteingress_protocol::*;
 
+/// Per-stream state tracked in the hub's stream map.
+struct HubStreamState {
+    /// Channel to deliver FRAME_DATA payloads to the upstream writer task.
+    data_tx: mpsc::Sender<Vec<u8>>,
+    /// Cancellation token for this stream.
+    cancel_token: CancellationToken,
+    /// Send window for FRAME_DATA_BACK (download direction).
+    /// Decremented by the upstream reader, incremented by FRAME_WINDOW_UPDATE from edge.
+    send_window: Arc<AtomicU32>,
+    /// Notifier to wake the upstream reader when the window opens.
+    window_notify: Arc<Notify>,
+}
+
 /// Hub configuration.
 #[derive(Debug, Clone, Deserialize, Serialize)]
 #[serde(rename_all = "camelCase")]
 pub struct HubConfig {
     pub tunnel_port: u16,
     pub target_host: Option<String>,
-    #[serde(skip)]
+    #[serde(default)]
     pub tls_cert_pem: Option<String>,
-    #[serde(skip)]
+    #[serde(default)]
     pub tls_key_pem: Option<String>,
 }
 
@@ -107,7 +123,7 @@ pub struct TunnelHub {
 struct ConnectedEdgeInfo {
     connected_at: u64,
     peer_addr: String,
-    active_streams: Arc<Mutex<HashMap<u32, (mpsc::Sender<Vec<u8>>, CancellationToken)>>>,
+    active_streams: Arc<Mutex<HashMap<u32, HubStreamState>>>,
     config_tx: mpsc::Sender<EdgeConfigUpdate>,
     #[allow(dead_code)] // kept alive for Drop — cancels child tokens when edge is removed
     cancel_token: CancellationToken,
@@ -331,7 +347,7 @@ async fn handle_edge_connection(
     write_half.write_all(handshake_json.as_bytes()).await?;
 
     // Track this edge
-    let streams: Arc<Mutex<HashMap<u32, (mpsc::Sender<Vec<u8>>, CancellationToken)>>> =
+    let streams: Arc<Mutex<HashMap<u32, HubStreamState>>> =
         Arc::new(Mutex::new(HashMap::new()));
     let now = std::time::SystemTime::now()
         .duration_since(std::time::UNIX_EPOCH)
@@ -355,14 +371,28 @@ async fn handle_edge_connection(
         );
     }
 
-    // A5: Channel-based writer replaces Arc<Mutex<WriteHalf>>
-    // All frame writes go through this channel → dedicated writer task serializes them
-    let (frame_writer_tx, mut frame_writer_rx) = mpsc::channel::<Vec<u8>>(4096);
+    // QoS dual-channel tunnel writer: control frames (PING/PONG/WINDOW_UPDATE/CLOSE)
+    // have priority over data frames (DATA_BACK). This prevents PING starvation under load.
+    let (ctrl_tx, mut ctrl_rx) = mpsc::channel::<Vec<u8>>(64);
+    let (data_tx, mut data_rx) = mpsc::channel::<Vec<u8>>(4096);
+    // Legacy alias for code that sends both control and data (will be migrated)
+    let frame_writer_tx = ctrl_tx.clone();
     let writer_token = edge_token.clone();
     let writer_handle = tokio::spawn(async move {
         loop {
             tokio::select! {
-                data = frame_writer_rx.recv() => {
+                biased; // control frames always take priority over data
+                ctrl = ctrl_rx.recv() => {
+                    match ctrl {
+                        Some(frame_data) => {
+                            if write_half.write_all(&frame_data).await.is_err() {
+                                break;
+                            }
+                        }
+                        None => break,
+                    }
+                }
+                data = data_rx.recv() => {
                     match data {
                         Some(frame_data) => {
                             if write_half.write_all(&frame_data).await.is_err() {
@@ -407,6 +437,14 @@ async fn handle_edge_connection(
     // A4: Semaphore to limit concurrent streams per edge
     let stream_semaphore = Arc::new(Semaphore::new(MAX_STREAMS_PER_EDGE));
 
+    // Heartbeat: periodic PING and liveness timeout
+    let ping_interval_dur = Duration::from_secs(15);
+    let liveness_timeout_dur = Duration::from_secs(45);
+    let mut ping_ticker = interval(ping_interval_dur);
+    ping_ticker.tick().await; // consume the immediate first tick
+    let mut last_activity = Instant::now();
+    let mut liveness_deadline = Box::pin(sleep_until(last_activity + liveness_timeout_dur));
+
     // Frame reading loop
     let mut frame_reader = FrameReader::new(buf_reader);
 
@@ -415,6 +453,10 @@ async fn handle_edge_connection(
             frame_result = frame_reader.next_frame() => {
                 match frame_result {
                     Ok(Some(frame)) => {
+                        // Reset liveness on any received frame
+                        last_activity = Instant::now();
+                        liveness_deadline.as_mut().reset(last_activity + liveness_timeout_dur);
+
                         match frame.frame_type {
                             FRAME_OPEN => {
                                 // A4: Check stream limit before processing
@@ -439,7 +481,8 @@ async fn handle_edge_connection(
                                 let edge_id_clone = edge_id.clone();
                                 let event_tx_clone = event_tx.clone();
                                 let streams_clone = streams.clone();
-                                let writer_tx = frame_writer_tx.clone();
+                                let writer_tx = ctrl_tx.clone();   // control: CLOSE_BACK, WINDOW_UPDATE_BACK
+                                let data_writer_tx = data_tx.clone(); // data: DATA_BACK
                                 let target = target_host.clone();
                                 let stream_token = edge_token.child_token();
 
@@ -448,11 +491,18 @@ async fn handle_edge_connection(
                                     stream_id,
                                 });
 
-                                // Create channel for data from edge to this stream
+                                // Create channel for data from edge to this stream (capacity 16 is sufficient with flow control)
                                 let (data_tx, mut data_rx) = mpsc::channel::<Vec<u8>>(256);
+                                let send_window = Arc::new(AtomicU32::new(INITIAL_STREAM_WINDOW));
+                                let window_notify = Arc::new(Notify::new());
                                 {
                                     let mut s = streams.lock().await;
-                                    s.insert(stream_id, (data_tx, stream_token.clone()));
+                                    s.insert(stream_id, HubStreamState {
+                                        data_tx,
+                                        cancel_token: stream_token.clone(),
+                                        send_window: Arc::clone(&send_window),
+                                        window_notify: Arc::clone(&window_notify),
+                                    });
                                 }
 
                                 // Spawn task: connect to SmartProxy, send PROXY header, pipe data
@@ -462,7 +512,7 @@ async fn handle_edge_connection(
                                     let result = async {
                                         // A2: Connect to SmartProxy with timeout
                                         let mut upstream = tokio::time::timeout(
-                                            std::time::Duration::from_secs(10),
+                                            Duration::from_secs(10),
                                             TcpStream::connect((target.as_str(), dest_port)),
                                         )
                                         .await
@@ -476,15 +526,36 @@ async fn handle_edge_connection(
                                             upstream.into_split();
 
                                         // Forward data from edge (via channel) to SmartProxy
+                                        // After writing to upstream, send WINDOW_UPDATE_BACK to edge
                                         let writer_token = stream_token.clone();
+                                        let wub_tx = writer_tx.clone();
                                         let writer_for_edge_data = tokio::spawn(async move {
+                                            let mut consumed_since_update: u32 = 0;
                                             loop {
                                                 tokio::select! {
                                                     data = data_rx.recv() => {
                                                         match data {
                                                             Some(data) => {
-                                                                if up_write.write_all(&data).await.is_err() {
-                                                                    break;
+                                                                let len = data.len() as u32;
+                                                                match tokio::time::timeout(
+                                                                    Duration::from_secs(60),
+                                                                    up_write.write_all(&data),
+                                                                ).await {
+                                                                    Ok(Ok(())) => {}
+                                                                    Ok(Err(_)) => break,
+                                                                    Err(_) => {
+                                                                        log::warn!("Stream {} write to upstream timed out (60s)", stream_id);
+                                                                        break;
+                                                                    }
+                                                                }
+                                                                // Track consumption for flow control
+                                                                consumed_since_update += len;
+                                                                if consumed_since_update >= WINDOW_UPDATE_THRESHOLD {
+                                                                    let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE_BACK, consumed_since_update);
+                                                                    if wub_tx.try_send(frame).is_ok() {
+                                                                        consumed_since_update = 0;
+                                                                    }
+                                                                    // If try_send fails, keep accumulating — retry on next threshold
                                                                 }
                                                             }
                                                             None => break,
@@ -493,22 +564,47 @@ async fn handle_edge_connection(
                                                     _ = writer_token.cancelled() => break,
                                                 }
                                             }
+                                            // Send final window update for remaining consumed bytes
+                                            if consumed_since_update > 0 {
+                                                let frame = encode_window_update(stream_id, FRAME_WINDOW_UPDATE_BACK, consumed_since_update);
+                                                let _ = wub_tx.try_send(frame);
+                                            }
                                             let _ = up_write.shutdown().await;
                                         });
 
                                         // Forward data from SmartProxy back to edge via writer channel
+                                        // with per-stream flow control (check send_window before reading)
                                         let mut buf = vec![0u8; 32768];
                                         loop {
+                                            // Wait for send window to have capacity (with stall timeout)
+                                            loop {
+                                                let w = send_window.load(Ordering::Acquire);
+                                                if w > 0 { break; }
+                                                tokio::select! {
+                                                    _ = window_notify.notified() => continue,
+                                                    _ = stream_token.cancelled() => break,
+                                                    _ = tokio::time::sleep(Duration::from_secs(120)) => {
+                                                        log::warn!("Stream {} download stalled (window empty for 120s)", stream_id);
+                                                        break;
+                                                    }
+                                                }
+                                            }
+                                            if stream_token.is_cancelled() { break; }
+
+                                            // Limit read size to available window
+                                            let w = send_window.load(Ordering::Acquire) as usize;
+                                            let max_read = w.min(buf.len());
+
                                             tokio::select! {
-                                                read_result = up_read.read(&mut buf) => {
+                                                read_result = up_read.read(&mut buf[..max_read]) => {
                                                     match read_result {
                                                         Ok(0) => break,
                                                         Ok(n) => {
+                                                            send_window.fetch_sub(n as u32, Ordering::Release);
                                                             let frame =
                                                                 encode_frame(stream_id, FRAME_DATA_BACK, &buf[..n]);
-                                                            // A5: Use try_send to avoid blocking if writer channel is full
-                                                            if writer_tx.try_send(frame).is_err() {
-                                                                log::warn!("Stream {} writer channel full, closing", stream_id);
+                                                            if data_writer_tx.send(frame).await.is_err() {
+                                                                log::warn!("Stream {} data channel closed, closing", stream_id);
                                                                 break;
                                                             }
                                                         }
@@ -519,10 +615,10 @@ async fn handle_edge_connection(
                                             }
                                         }
 
-                                        // Send CLOSE_BACK to edge (only if not cancelled)
+                                        // Send CLOSE_BACK via DATA channel (must arrive AFTER last DATA_BACK)
                                         if !stream_token.is_cancelled() {
                                             let close_frame = encode_frame(stream_id, FRAME_CLOSE_BACK, &[]);
-                                            let _ = writer_tx.try_send(close_frame);
+                                            let _ = data_writer_tx.try_send(close_frame);
                                         }
 
                                         writer_for_edge_data.abort();
@@ -532,10 +628,10 @@ async fn handle_edge_connection(
 
                                     if let Err(e) = result {
                                         log::error!("Stream {} error: {}", stream_id, e);
-                                        // Send CLOSE_BACK on error (only if not cancelled)
+                                        // Send CLOSE_BACK via DATA channel on error (must arrive after any DATA_BACK)
                                         if !stream_token.is_cancelled() {
                                             let close_frame = encode_frame(stream_id, FRAME_CLOSE_BACK, &[]);
-                                            let _ = writer_tx.try_send(close_frame);
+                                            let _ = data_writer_tx.try_send(close_frame);
                                         }
                                     }
 
@@ -553,24 +649,46 @@ async fn handle_edge_connection(
                                 });
                             }
                             FRAME_DATA => {
-                                // A1: Non-blocking send to prevent head-of-line blocking
-                                let s = streams.lock().await;
-                                if let Some((tx, _)) = s.get(&frame.stream_id) {
-                                    if tx.try_send(frame.payload).is_err() {
-                                        log::warn!("Stream {} data channel full, dropping frame", frame.stream_id);
+                                // Non-blocking dispatch to per-stream channel.
+                                // With flow control, the sender should rarely exceed the channel capacity.
+                                let mut s = streams.lock().await;
+                                if let Some(state) = s.get(&frame.stream_id) {
+                                    if state.data_tx.try_send(frame.payload).is_err() {
+                                        log::warn!("Stream {} data channel full, closing stream", frame.stream_id);
+                                        if let Some(state) = s.remove(&frame.stream_id) {
+                                            state.cancel_token.cancel();
+                                        }
+                                    }
+                                }
+                            }
+                            FRAME_WINDOW_UPDATE => {
+                                // Edge consumed data — increase our send window for this stream
+                                if let Some(increment) = decode_window_update(&frame.payload) {
+                                    if increment > 0 {
+                                        let s = streams.lock().await;
+                                        if let Some(state) = s.get(&frame.stream_id) {
+                                            let prev = state.send_window.fetch_add(increment, Ordering::Release);
+                                            if prev + increment > MAX_WINDOW_SIZE {
+                                                state.send_window.store(MAX_WINDOW_SIZE, Ordering::Release);
+                                            }
+                                            state.window_notify.notify_one();
+                                        }
                                     }
                                 }
                             }
                             FRAME_CLOSE => {
                                 let mut s = streams.lock().await;
-                                if let Some((_, token)) = s.remove(&frame.stream_id) {
-                                    token.cancel();
+                                if let Some(state) = s.remove(&frame.stream_id) {
+                                    state.cancel_token.cancel();
                                     let _ = event_tx.try_send(HubEvent::StreamClosed {
                                         edge_id: edge_id.clone(),
                                         stream_id: frame.stream_id,
                                     });
                                 }
                             }
+                            FRAME_PONG => {
+                                log::debug!("Received PONG from edge {}", edge_id);
+                            }
                             _ => {
                                 log::warn!("Unexpected frame type {} from edge", frame.frame_type);
                             }
@@ -586,6 +704,19 @@ async fn handle_edge_connection(
                     }
                 }
             }
+            _ = ping_ticker.tick() => {
+                let ping_frame = encode_frame(0, FRAME_PING, &[]);
+                if frame_writer_tx.try_send(ping_frame).is_err() {
+                    log::warn!("Failed to send PING to edge {}, writer channel full/closed", edge_id);
+                    break;
+                }
+                log::trace!("Sent PING to edge {}", edge_id);
+            }
+            _ = &mut liveness_deadline => {
+                log::warn!("Edge {} liveness timeout (no frames for {}s), disconnecting",
+                    edge_id, liveness_timeout_dur.as_secs());
+                break;
+            }
             _ = edge_token.cancelled() => {
                 log::info!("Edge {} cancelled by hub", edge_id);
                 break;

rust/crates/remoteingress-protocol/src/lib.rs

@@ -7,6 +7,10 @@ pub const FRAME_CLOSE: u8 = 0x03;
 pub const FRAME_DATA_BACK: u8 = 0x04;
 pub const FRAME_CLOSE_BACK: u8 = 0x05;
 pub const FRAME_CONFIG: u8 = 0x06;     // Hub -> Edge: configuration update
+pub const FRAME_PING: u8 = 0x07;       // Hub -> Edge: heartbeat probe
+pub const FRAME_PONG: u8 = 0x08;       // Edge -> Hub: heartbeat response
+pub const FRAME_WINDOW_UPDATE: u8 = 0x09;      // Edge -> Hub: per-stream flow control
+pub const FRAME_WINDOW_UPDATE_BACK: u8 = 0x0A;  // Hub -> Edge: per-stream flow control
 
 // Frame header size: 4 (stream_id) + 1 (type) + 4 (length) = 9 bytes
 pub const FRAME_HEADER_SIZE: usize = 9;
@@ -14,6 +18,28 @@ pub const FRAME_HEADER_SIZE: usize = 9;
 // Maximum payload size (16 MB)
 pub const MAX_PAYLOAD_SIZE: u32 = 16 * 1024 * 1024;
 
+// Per-stream flow control constants
+/// Initial per-stream window size (4 MB). Sized for full throughput at high RTT:
+/// at 100ms RTT, this sustains ~40 MB/s per stream.
+pub const INITIAL_STREAM_WINDOW: u32 = 4 * 1024 * 1024;
+/// Send WINDOW_UPDATE after consuming this many bytes (half the initial window).
+pub const WINDOW_UPDATE_THRESHOLD: u32 = INITIAL_STREAM_WINDOW / 2;
+/// Maximum window size to prevent overflow.
+pub const MAX_WINDOW_SIZE: u32 = 16 * 1024 * 1024;
+
+/// Encode a WINDOW_UPDATE frame for a specific stream.
+pub fn encode_window_update(stream_id: u32, frame_type: u8, increment: u32) -> Vec<u8> {
+    encode_frame(stream_id, frame_type, &increment.to_be_bytes())
+}
+
+/// Decode a WINDOW_UPDATE payload into a byte increment. Returns None if payload is malformed.
+pub fn decode_window_update(payload: &[u8]) -> Option<u32> {
+    if payload.len() != 4 {
+        return None;
+    }
+    Some(u32::from_be_bytes([payload[0], payload[1], payload[2], payload[3]]))
+}
+
 /// A single multiplexed frame.
 #[derive(Debug, Clone)]
 pub struct Frame {
@@ -261,6 +287,8 @@ mod tests {
             FRAME_DATA_BACK,
             FRAME_CLOSE_BACK,
             FRAME_CONFIG,
+            FRAME_PING,
+            FRAME_PONG,
         ];
 
         let mut data = Vec::new();
@@ -293,4 +321,19 @@ mod tests {
         assert_eq!(frame.frame_type, FRAME_CLOSE);
         assert!(frame.payload.is_empty());
     }
+
+    #[test]
+    fn test_encode_frame_ping_pong() {
+        // PING: stream_id=0, empty payload (control frame)
+        let ping = encode_frame(0, FRAME_PING, &[]);
+        assert_eq!(ping[4], FRAME_PING);
+        assert_eq!(&ping[0..4], &0u32.to_be_bytes());
+        assert_eq!(ping.len(), FRAME_HEADER_SIZE);
+
+        // PONG: stream_id=0, empty payload (control frame)
+        let pong = encode_frame(0, FRAME_PONG, &[]);
+        assert_eq!(pong[4], FRAME_PONG);
+        assert_eq!(&pong[0..4], &0u32.to_be_bytes());
+        assert_eq!(pong.len(), FRAME_HEADER_SIZE);
+    }
 }

ts/00_commitinfo_data.ts

@@ -3,6 +3,6 @@
  */
 export const commitinfo = {
   name: '@serve.zone/remoteingress',
-  version: '4.2.0',
+  version: '4.5.5',
   description: 'Edge ingress tunnel for DcRouter - accepts incoming TCP connections at network edge and tunnels them to DcRouter SmartProxy preserving client IP via PROXY protocol v1.'
 }

ts/classes.remoteingressedge.ts

@@ -40,9 +40,16 @@ export interface IEdgeConfig {
   secret: string;
 }
 
+const MAX_RESTART_ATTEMPTS = 10;
+const MAX_RESTART_BACKOFF_MS = 30_000;
+
 export class RemoteIngressEdge extends EventEmitter {
   private bridge: InstanceType<typeof plugins.smartrust.RustBridge<TEdgeCommands>>;
   private started = false;
+  private stopping = false;
+  private savedConfig: IEdgeConfig | null = null;
+  private restartBackoffMs = 1000;
+  private restartAttempts = 0;
   private statusInterval: ReturnType<typeof setInterval> | undefined;
 
   constructor() {
@@ -109,11 +116,17 @@ export class RemoteIngressEdge extends EventEmitter {
       edgeConfig = config;
     }
 
+    this.savedConfig = edgeConfig;
+    this.stopping = false;
+
     const spawned = await this.bridge.spawn();
     if (!spawned) {
       throw new Error('Failed to spawn remoteingress-bin');
     }
 
+    // Register crash recovery handler
+    this.bridge.on('exit', this.handleCrashRecovery);
+
     await this.bridge.sendCommand('startEdge', {
       hubHost: edgeConfig.hubHost,
       hubPort: edgeConfig.hubPort ?? 8443,
@@ -122,6 +135,8 @@ export class RemoteIngressEdge extends EventEmitter {
     });
 
     this.started = true;
+    this.restartAttempts = 0;
+    this.restartBackoffMs = 1000;
 
     // Start periodic status logging
     this.statusInterval = setInterval(async () => {
@@ -142,6 +157,7 @@ export class RemoteIngressEdge extends EventEmitter {
    * Stop the edge and kill the Rust process.
    */
   public async stop(): Promise<void> {
+    this.stopping = true;
     if (this.statusInterval) {
       clearInterval(this.statusInterval);
       this.statusInterval = undefined;
@@ -152,6 +168,7 @@ export class RemoteIngressEdge extends EventEmitter {
       } catch {
         // Process may already be dead
       }
+      this.bridge.removeListener('exit', this.handleCrashRecovery);
       this.bridge.kill();
       this.started = false;
     }
@@ -170,4 +187,55 @@ export class RemoteIngressEdge extends EventEmitter {
   public get running(): boolean {
     return this.bridge.running;
   }
+
+  /**
+   * Handle unexpected Rust binary crash — auto-restart with backoff.
+   */
+  private handleCrashRecovery = async (code: number | null, signal: string | null) => {
+    if (this.stopping || !this.started || !this.savedConfig) {
+      return;
+    }
+
+    console.error(
+      `[RemoteIngressEdge] Rust binary crashed (code=${code}, signal=${signal}), ` +
+      `attempt ${this.restartAttempts + 1}/${MAX_RESTART_ATTEMPTS}`
+    );
+
+    this.started = false;
+
+    if (this.restartAttempts >= MAX_RESTART_ATTEMPTS) {
+      console.error('[RemoteIngressEdge] Max restart attempts reached, giving up');
+      this.emit('crashRecoveryFailed');
+      return;
+    }
+
+    await new Promise(resolve => setTimeout(resolve, this.restartBackoffMs));
+    this.restartBackoffMs = Math.min(this.restartBackoffMs * 2, MAX_RESTART_BACKOFF_MS);
+    this.restartAttempts++;
+
+    try {
+      const spawned = await this.bridge.spawn();
+      if (!spawned) {
+        console.error('[RemoteIngressEdge] Failed to respawn binary');
+        return;
+      }
+
+      this.bridge.on('exit', this.handleCrashRecovery);
+
+      await this.bridge.sendCommand('startEdge', {
+        hubHost: this.savedConfig.hubHost,
+        hubPort: this.savedConfig.hubPort ?? 8443,
+        edgeId: this.savedConfig.edgeId,
+        secret: this.savedConfig.secret,
+      });
+
+      this.started = true;
+      this.restartAttempts = 0;
+      this.restartBackoffMs = 1000;
+      console.log('[RemoteIngressEdge] Successfully recovered from crash');
+      this.emit('crashRecovered');
+    } catch (err) {
+      console.error(`[RemoteIngressEdge] Crash recovery failed: ${err}`);
+    }
+  };
 }

ts/classes.remoteingresshub.ts

@@ -11,6 +11,8 @@ type THubCommands = {
     params: {
       tunnelPort: number;
       targetHost?: string;
+      tlsCertPem?: string;
+      tlsKeyPem?: string;
     };
     result: { started: boolean };
   };
@@ -42,11 +44,25 @@ type THubCommands = {
 export interface IHubConfig {
   tunnelPort?: number;
   targetHost?: string;
+  tls?: {
+    certPem?: string;
+    keyPem?: string;
+  };
 }
 
+type TAllowedEdge = { id: string; secret: string; listenPorts?: number[]; stunIntervalSecs?: number };
+
+const MAX_RESTART_ATTEMPTS = 10;
+const MAX_RESTART_BACKOFF_MS = 30_000;
+
 export class RemoteIngressHub extends EventEmitter {
   private bridge: InstanceType<typeof plugins.smartrust.RustBridge<THubCommands>>;
   private started = false;
+  private stopping = false;
+  private savedConfig: IHubConfig | null = null;
+  private savedEdges: TAllowedEdge[] = [];
+  private restartBackoffMs = 1000;
+  private restartAttempts = 0;
 
   constructor() {
     super();
@@ -92,29 +108,42 @@ export class RemoteIngressHub extends EventEmitter {
    * Start the hub — spawns the Rust binary and starts the tunnel server.
    */
   public async start(config: IHubConfig = {}): Promise<void> {
+    this.savedConfig = config;
+    this.stopping = false;
+
     const spawned = await this.bridge.spawn();
     if (!spawned) {
       throw new Error('Failed to spawn remoteingress-bin');
     }
 
+    // Register crash recovery handler
+    this.bridge.on('exit', this.handleCrashRecovery);
+
     await this.bridge.sendCommand('startHub', {
       tunnelPort: config.tunnelPort ?? 8443,
       targetHost: config.targetHost ?? '127.0.0.1',
+      ...(config.tls?.certPem && config.tls?.keyPem
+        ? { tlsCertPem: config.tls.certPem, tlsKeyPem: config.tls.keyPem }
+        : {}),
     });
 
     this.started = true;
+    this.restartAttempts = 0;
+    this.restartBackoffMs = 1000;
   }
 
   /**
    * Stop the hub and kill the Rust process.
    */
   public async stop(): Promise<void> {
+    this.stopping = true;
     if (this.started) {
       try {
         await this.bridge.sendCommand('stopHub', {} as Record<string, never>);
       } catch {
         // Process may already be dead
       }
+      this.bridge.removeListener('exit', this.handleCrashRecovery);
       this.bridge.kill();
       this.started = false;
     }
@@ -123,7 +152,8 @@ export class RemoteIngressHub extends EventEmitter {
   /**
    * Update the list of allowed edges that can connect to this hub.
    */
-  public async updateAllowedEdges(edges: Array<{ id: string; secret: string; listenPorts?: number[]; stunIntervalSecs?: number }>): Promise<void> {
+  public async updateAllowedEdges(edges: TAllowedEdge[]): Promise<void> {
+    this.savedEdges = edges;
     await this.bridge.sendCommand('updateAllowedEdges', { edges });
   }
 
@@ -140,4 +170,62 @@ export class RemoteIngressHub extends EventEmitter {
   public get running(): boolean {
     return this.bridge.running;
   }
+
+  /**
+   * Handle unexpected Rust binary crash — auto-restart with backoff.
+   */
+  private handleCrashRecovery = async (code: number | null, signal: string | null) => {
+    if (this.stopping || !this.started || !this.savedConfig) {
+      return;
+    }
+
+    console.error(
+      `[RemoteIngressHub] Rust binary crashed (code=${code}, signal=${signal}), ` +
+      `attempt ${this.restartAttempts + 1}/${MAX_RESTART_ATTEMPTS}`
+    );
+
+    this.started = false;
+
+    if (this.restartAttempts >= MAX_RESTART_ATTEMPTS) {
+      console.error('[RemoteIngressHub] Max restart attempts reached, giving up');
+      this.emit('crashRecoveryFailed');
+      return;
+    }
+
+    await new Promise(resolve => setTimeout(resolve, this.restartBackoffMs));
+    this.restartBackoffMs = Math.min(this.restartBackoffMs * 2, MAX_RESTART_BACKOFF_MS);
+    this.restartAttempts++;
+
+    try {
+      const spawned = await this.bridge.spawn();
+      if (!spawned) {
+        console.error('[RemoteIngressHub] Failed to respawn binary');
+        return;
+      }
+
+      this.bridge.on('exit', this.handleCrashRecovery);
+
+      const config = this.savedConfig;
+      await this.bridge.sendCommand('startHub', {
+        tunnelPort: config.tunnelPort ?? 8443,
+        targetHost: config.targetHost ?? '127.0.0.1',
+        ...(config.tls?.certPem && config.tls?.keyPem
+          ? { tlsCertPem: config.tls.certPem, tlsKeyPem: config.tls.keyPem }
+          : {}),
+      });
+
+      // Restore allowed edges
+      if (this.savedEdges.length > 0) {
+        await this.bridge.sendCommand('updateAllowedEdges', { edges: this.savedEdges });
+      }
+
+      this.started = true;
+      this.restartAttempts = 0;
+      this.restartBackoffMs = 1000;
+      console.log('[RemoteIngressHub] Successfully recovered from crash');
+      this.emit('crashRecovered');
+    } catch (err) {
+      console.error(`[RemoteIngressHub] Crash recovery failed: ${err}`);
+    }
+  };
 }