Harden alternate screen tracking

Cargo.lock

@@ -436,7 +436,7 @@ dependencies = [
 
 [[package]]
 name = "dosh"
-version = "0.1.10"
+version = "1.0.0-rc9"
 dependencies = [
  "anyhow",
  "base64",

Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "dosh"
-version = "0.1.10"
+version = "1.0.0-rc9"
 edition = "2024"
 license = "MIT"
 

Makefile

@@ -1,4 +1,4 @@
-.PHONY: build test fmt clippy release-check install package-release package-release-linux package-release-windows publish-release bench-report bench-local bench-local-json bench-docker-ssh bench-docker-mosh fuzz-smoke fuzz-deep soak-local tui-harness
+.PHONY: build test fmt clippy release-check 1.0-check reconnect-check hostile-network-check persistence-check package-check install package-release package-release-linux package-release-windows publish-release bench-report bench-local bench-local-json bench-docker-ssh bench-docker-mosh fuzz-smoke fuzz-deep soak-local tui-harness
 
 build:
 	cargo build --release
@@ -19,6 +19,25 @@ release-check:
 	cargo test
 	$(MAKE) tui-harness
 
+1.0-check: release-check reconnect-check hostile-network-check persistence-check package-check
+
+reconnect-check:
+	cargo test --test integration_smoke resume_updates_udp_endpoint_for_roaming -- --nocapture
+	DOSH_SOAK_SECONDS=$${DOSH_SOAK_SECONDS:-300} cargo test --test integration_smoke sleep_roaming_soak_30m -- --ignored --nocapture
+
+hostile-network-check:
+	cargo test --test hostile_network -- --nocapture
+	DOSH_BADNET_SOAK_SECONDS=$${DOSH_BADNET_SOAK_SECONDS:-300} cargo test --test hostile_network bad_network_tui_work_soak_30m -- --ignored --nocapture
+
+persistence-check:
+	cargo test --test integration_smoke session_survives_server_restart_same_shell_and_screen -- --nocapture
+	cargo test --test integration_smoke multiple_persistent_named_sessions_survive_restart_independently -- --nocapture
+
+package-check:
+	$(MAKE) package-release-linux
+	$(MAKE) package-release-windows
+	sh scripts/verify-release-artifacts.sh
+
 install:
 	sh install.sh --from-current
 

install.sh

@@ -426,18 +426,9 @@ Wants=network-online.target
 [Service]
 Type=simple
 ExecStart=$bindir/dosh-server serve
-Restart=on-failure
+Restart=always
 RestartSec=1
 KillMode=process
-NoNewPrivileges=true
-PrivateTmp=true
-ProtectSystem=strict
-ProtectHome=read-only
-ReadWritePaths=$config_dir $data_dir
-RestrictAddressFamilies=AF_INET AF_INET6 AF_UNIX
-RestrictRealtime=true
-LockPersonality=true
-MemoryDenyWriteExecute=true
 
 [Install]
 WantedBy=default.target

packaging/systemd/dosh-server.service

@@ -6,18 +6,9 @@ Wants=network-online.target
 [Service]
 Type=simple
 ExecStart=%h/.local/bin/dosh-server serve
-Restart=on-failure
+Restart=always
 RestartSec=1
 KillMode=process
-NoNewPrivileges=true
-PrivateTmp=true
-ProtectSystem=strict
-ProtectHome=read-only
-ReadWritePaths=%h/.config/dosh %h/.local/share/dosh
-RestrictAddressFamilies=AF_INET AF_INET6 AF_UNIX
-RestrictRealtime=true
-LockPersonality=true
-MemoryDenyWriteExecute=true
 
 [Install]
 WantedBy=default.target

scripts/bench-local.sh

@@ -67,6 +67,7 @@ cleanup() {
     kill "$server_pid" 2>/dev/null || true
     wait "$server_pid" 2>/dev/null || true
   fi
+  pkill -f "$server_bin hold --runtime-dir $home/sessions/run" 2>/dev/null || true
   rm -rf "$home"
 }
 trap cleanup EXIT INT TERM
@@ -86,6 +87,7 @@ scrollback = 5000
 auth_ttl_secs = 30
 attach_ticket_ttl_secs = 3600
 allow_attach_tickets = true
+persist_sessions = false
 client_timeout_secs = 60
 retransmit_window = 256
 default_input_mode = "read-write"

scripts/verify-release-artifacts.sh

@@ -0,0 +1,46 @@
+#!/usr/bin/env sh
+set -eu
+
+repo_root="$(CDPATH= cd -- "$(dirname -- "$0")/.." && pwd)"
+cd "$repo_root"
+
+version="$(sed -n 's/^version = "\(.*\)"/\1/p' Cargo.toml | sed -n '1p')"
+out_dir="${DOSH_PACKAGE_DIR:-target/dosh-release}"
+
+artifact_version() {
+  artifact="$1"
+  case "$artifact" in
+    *.tar.gz)
+      tar -xOf "$artifact" dosh/VERSION 2>/dev/null | tr -d '\r\n'
+      ;;
+    *.zip)
+      unzip -p "$artifact" dosh/VERSION 2>/dev/null | tr -d '\r\n'
+      ;;
+    *)
+      return 1
+      ;;
+  esac
+}
+
+failed=0
+for artifact in \
+  "$out_dir/dosh-linux-x86_64.tar.gz" \
+  "$out_dir/dosh-macos-aarch64.tar.gz" \
+  "$out_dir/dosh-windows-x86_64.zip"; do
+  if [ ! -f "$artifact" ]; then
+    echo "missing release artifact: $artifact" >&2
+    failed=1
+    continue
+  fi
+  if [ ! -f "$artifact.sha256" ]; then
+    echo "missing release checksum: $artifact.sha256" >&2
+    failed=1
+  fi
+  actual="$(artifact_version "$artifact" || true)"
+  if [ "$actual" != "$version" ]; then
+    echo "artifact version mismatch: $artifact has ${actual:-unknown}, expected $version" >&2
+    failed=1
+  fi
+done
+
+exit "$failed"

src/bin/dosh-server.rs

@@ -28,6 +28,7 @@ use dosh::protocol::{
     TicketAttachBody, TicketAttachEnvelope, TicketAttachOkEnvelope,
 };
 use dosh::pty::{PtyHandle, PtyOutput, adopt_pty_from_fd, spawn_pty_session};
+use dosh::udp::is_transient_udp_send_error;
 use sha2::{Digest, Sha256};
 use std::collections::{BTreeMap, HashMap, HashSet, VecDeque};
 use std::fs;
@@ -51,6 +52,7 @@ const STREAM_INITIAL_WINDOW: usize = 1024 * 1024;
 /// have accumulated since the last mirror. Keeps the atomic write off the
 /// per-packet hot path while bounding how much fresh output a crash can lose.
 const SCREEN_PERSIST_BYTES: usize = 4096;
+const IMPLICIT_EMPTY_SESSION_GRACE_SECS: u64 = 300;
 
 /// Sentinel `target_host` sent to the client on a server-initiated `StreamOpen`
 /// that carries an SSH-agent connection (the client splices it into its local
@@ -432,6 +434,7 @@ struct PendingStreamOpen {
     attempts: u32,
 }
 
+#[derive(Clone)]
 struct PendingNativeAuth {
     client: dosh::native::NativeClientHello,
     server: NativeServerHello,
@@ -510,12 +513,14 @@ impl ServerState {
         Ok(())
     }
 
-    /// Whether a session named `name` should be backed by a persistent holder.
+    /// Whether a session should be backed by a persistent holder.
     ///
-    /// With persistence enabled, every PTY session gets a detached holder. That
-    /// includes implicit `term-<millis>-<pid>` sessions: a reconnecting client has
-    /// the exact session name in its ticket cache, so it can reattach after a
-    /// quick server restart without forcing users to name sessions manually.
+    /// With persistence enabled, every interactive session gets a detached holder
+    /// so an update/restart of `dosh-server` behaves like a transient network
+    /// break. Named sessions can sit empty for the normal client timeout.
+    /// Generated one-off sessions are also persisted, but cleanup gives them a
+    /// shorter empty grace period so invisible abandoned holders do not linger
+    /// indefinitely.
     fn should_persist_session(&self, _name: &str) -> bool {
         self.config.persist_sessions
     }
@@ -654,6 +659,7 @@ impl ServerState {
     fn insert_client(&mut self, session_name: &str, client_id: [u8; 16], client: ClientState) {
         if let Some(session) = self.sessions.get_mut(session_name) {
             session.clients.insert(client_id, client);
+            session.empty_since = None;
             self.client_index
                 .insert(client_id, session_name.to_string());
         }
@@ -848,6 +854,7 @@ async fn handle_packet(
         | PacketKind::StreamEof
         | PacketKind::StreamWindowAdjust
         | PacketKind::RekeyAck
+        | PacketKind::Detach
             if find_client_decrypt_key(state, &packet.header).is_err() =>
         {
             send_reject_to_client(socket, peer, packet.header.conn_id, "unknown client").await
@@ -965,7 +972,7 @@ async fn handle_native_client_hello(
     };
     let body = protocol::to_body(&NativeServerHelloBody { hello })?;
     let out = protocol::encode_plain(PacketKind::NativeServerHello, pending_id, 1, 0, &body)?;
-    socket.send_to(&out, peer).await?;
+    let _ = send_udp(socket, &out, peer).await?;
     Ok(())
 }
 
@@ -976,8 +983,8 @@ async fn handle_native_user_auth(
     packet: &protocol::Packet,
 ) -> Result<()> {
     let pending = {
-        let mut locked = state.lock().expect("server state poisoned");
-        locked.pending_native.remove(&packet.header.conn_id)
+        let locked = state.lock().expect("server state poisoned");
+        locked.pending_native.get(&packet.header.conn_id).cloned()
     };
     let Some(pending) = pending else {
         return send_reject_to_client(socket, peer, packet.header.conn_id, "unknown native auth")
@@ -993,11 +1000,38 @@ async fn handle_native_user_auth(
         .await;
     }
     if pending.created_at.elapsed() > Duration::from_secs(30) {
+        state
+            .lock()
+            .expect("server state poisoned")
+            .pending_native
+            .remove(&packet.header.conn_id);
         return send_reject_to_client(socket, peer, packet.header.conn_id, "native auth expired")
             .await;
     }
-    let body = protocol::decrypt_body(packet, &pending.session_key, CLIENT_TO_SERVER)?;
-    let req: NativeUserAuthBody = protocol::from_body(&body)?;
+    let body = match protocol::decrypt_body(packet, &pending.session_key, CLIENT_TO_SERVER) {
+        Ok(body) => body,
+        Err(_) => {
+            return send_reject_to_client(
+                socket,
+                peer,
+                packet.header.conn_id,
+                "native auth decrypt failed",
+            )
+            .await;
+        }
+    };
+    let req: NativeUserAuthBody = match protocol::from_body(&body) {
+        Ok(req) => req,
+        Err(_) => {
+            return send_reject_to_client(
+                socket,
+                peer,
+                packet.header.conn_id,
+                "invalid native auth body",
+            )
+            .await;
+        }
+    };
     if pending.client.requested_mode == "doctor" {
         let check_result = {
             let locked = state.lock().expect("server state poisoned");
@@ -1033,6 +1067,11 @@ async fn handle_native_user_auth(
                 .await;
             }
         };
+        state
+            .lock()
+            .expect("server state poisoned")
+            .pending_native
+            .remove(&packet.header.conn_id);
         let body = protocol::to_body(&check)?;
         let out = protocol::encode_encrypted(
             PacketKind::NativeAuthCheckOk,
@@ -1043,7 +1082,7 @@ async fn handle_native_user_auth(
             SERVER_TO_CLIENT,
             &body,
         )?;
-        socket.send_to(&out, peer).await?;
+        let _ = send_udp(socket, &out, peer).await?;
         return Ok(());
     }
 
@@ -1213,6 +1252,11 @@ async fn handle_native_user_auth(
     if let Some((listener, path)) = agent_listener {
         spawn_agent_forward(state, socket, client_id, listener, path);
     }
+    state
+        .lock()
+        .expect("server state poisoned")
+        .pending_native
+        .remove(&packet.header.conn_id);
 
     let ok = NativeAuthOk {
         client_id,
@@ -1236,7 +1280,7 @@ async fn handle_native_user_auth(
         SERVER_TO_CLIENT,
         &body,
     )?;
-    socket.send_to(&out, peer).await?;
+    let _ = send_udp(socket, &out, peer).await?;
     Ok(())
 }
 
@@ -1344,7 +1388,7 @@ async fn handle_bootstrap_attach(
         SERVER_TO_CLIENT,
         &body,
     )?;
-    socket.send_to(&out, peer).await?;
+    let _ = send_udp(socket, &out, peer).await?;
     Ok(())
 }
 
@@ -1479,7 +1523,7 @@ async fn handle_ticket_attach(
     };
     let body = protocol::to_body(&envelope)?;
     let out = protocol::encode_plain(PacketKind::AttachOk, client_id, 1, 0, &body)?;
-    socket.send_to(&out, peer).await?;
+    let _ = send_udp(socket, &out, peer).await?;
     Ok(())
 }
 
@@ -1497,7 +1541,7 @@ async fn send_reject_to_client(
         reason: reason.to_string(),
     })?;
     let out = protocol::encode_plain(PacketKind::AttachReject, client_id, 0, 0, &body)?;
-    socket.send_to(&out, peer).await?;
+    let _ = send_udp(socket, &out, peer).await?;
     Ok(())
 }
 
@@ -1563,7 +1607,7 @@ async fn handle_resume(
         SERVER_TO_CLIENT,
         &body,
     )?;
-    socket.send_to(&out, peer).await?;
+    let _ = send_udp(socket, &out, peer).await?;
     Ok(())
 }
 
@@ -1675,12 +1719,19 @@ async fn handle_ping(
         SERVER_TO_CLIENT,
         b"",
     )?;
-    socket.send_to(&out, peer).await?;
+    let _ = send_udp(socket, &out, peer).await?;
     Ok(())
 }
 
 async fn handle_detach(state: &Arc<Mutex<ServerState>>, packet: &protocol::Packet) -> Result<()> {
+    let (key, _) = find_client_decrypt_key(state, &packet.header)?;
+    let _ = protocol::decrypt_body(packet, &key, CLIENT_TO_SERVER)?;
     let mut locked = state.lock().expect("server state poisoned");
+    if let Some(client) = locked.client_mut(&packet.header.conn_id) {
+        if !client.replay.accept(packet.header.seq) {
+            return Ok(());
+        }
+    }
     locked.remove_client_everywhere(&packet.header.conn_id);
     Ok(())
 }
@@ -1690,6 +1741,14 @@ fn remove_client(state: &Arc<Mutex<ServerState>>, client_id: [u8; 16]) {
     locked.remove_client_everywhere(&client_id);
 }
 
+async fn send_udp(socket: &UdpSocket, packet: &[u8], peer: SocketAddr) -> Result<bool> {
+    match socket.send_to(packet, peer).await {
+        Ok(_) => Ok(true),
+        Err(err) if is_transient_udp_send_error(&err) => Ok(false),
+        Err(err) => Err(err).with_context(|| format!("send UDP packet to {peer}")),
+    }
+}
+
 async fn handle_ack(
     state: &Arc<Mutex<ServerState>>,
     peer: SocketAddr,
@@ -3005,7 +3064,7 @@ async fn send_stream_open_to_client(
         found
     };
     if let Some((endpoint, packet)) = send {
-        socket.send_to(&packet, endpoint).await?;
+        let _ = send_udp(socket, &packet, endpoint).await?;
     }
     Ok(())
 }
@@ -3099,7 +3158,7 @@ async fn queue_or_send_stream_data_to_client(
         send
     };
     if let Some((endpoint, packet)) = send {
-        socket.send_to(&packet, endpoint).await?;
+        let _ = send_udp(socket, &packet, endpoint).await?;
     }
     Ok(())
 }
@@ -3174,7 +3233,7 @@ async fn flush_stream_pending_data_to_client(
         let Some((endpoint, packet)) = send else {
             return Ok(());
         };
-        socket.send_to(&packet, endpoint).await?;
+        let _ = send_udp(socket, &packet, endpoint).await?;
     }
 }
 
@@ -3318,7 +3377,7 @@ async fn send_stream_packet_to_client(
         found
     };
     if let Some((endpoint, packet)) = send {
-        socket.send_to(&packet, endpoint).await?;
+        let _ = send_udp(socket, &packet, endpoint).await?;
     }
     Ok(())
 }
@@ -3420,7 +3479,7 @@ async fn broadcast_output(
         }
     }
     for (endpoint, packet) in sends {
-        socket.send_to(&packet, endpoint).await?;
+        let _ = send_udp(socket, &packet, endpoint).await?;
     }
     Ok(())
 }
@@ -3473,7 +3532,7 @@ async fn broadcast_session_exit(
         persist::remove_runtime_dir(&sessions_dir, &session_name);
     }
     for (endpoint, packet) in sends {
-        socket.send_to(&packet, endpoint).await?;
+        let _ = send_udp(socket, &packet, endpoint).await?;
     }
     Ok(())
 }
@@ -3504,11 +3563,9 @@ async fn retransmit_pending(
                     if pending.output_seq <= client.last_acked {
                         continue;
                     }
-                    if now.duration_since(pending.last_sent) >= Duration::from_millis(200)
-                        && pending.attempts < 8
-                    {
+                    if now.duration_since(pending.last_sent) >= Duration::from_millis(200) {
                         pending.last_sent = now;
-                        pending.attempts += 1;
+                        pending.attempts = pending.attempts.saturating_add(1);
                         sends.push((client.endpoint, pending.packet.clone()));
                     }
                 }
@@ -3582,7 +3639,7 @@ async fn retransmit_pending(
         sends
     };
     for (endpoint, packet) in sends {
-        socket.send_to(&packet, endpoint).await?;
+        let _ = send_udp(socket, &packet, endpoint).await?;
     }
     Ok(())
 }
@@ -3678,7 +3735,7 @@ async fn maybe_rekey_clients(
         sends
     };
     for (endpoint, packet) in sends {
-        socket.send_to(&packet, endpoint).await?;
+        let _ = send_udp(socket, &packet, endpoint).await?;
     }
     Ok(())
 }
@@ -3774,9 +3831,9 @@ fn cleanup_disconnected_clients(state: &Arc<Mutex<ServerState>>) {
             .iter()
             .filter(|(name, session)| {
                 !prewarm.contains(name.as_str())
-                    && session
-                        .empty_since
-                        .is_some_and(|since| now.duration_since(since) >= timeout)
+                    && session.empty_since.is_some_and(|since| {
+                        now.duration_since(since) >= empty_session_timeout(name, timeout)
+                    })
             })
             .map(|(name, _)| name.clone())
             .collect();
@@ -3795,6 +3852,14 @@ fn cleanup_disconnected_clients(state: &Arc<Mutex<ServerState>>) {
     }
 }
 
+fn empty_session_timeout(name: &str, configured_timeout: Duration) -> Duration {
+    if protocol::is_implicit_session_name(name) {
+        configured_timeout.min(Duration::from_secs(IMPLICIT_EMPTY_SESSION_GRACE_SECS))
+    } else {
+        configured_timeout
+    }
+}
+
 /// Select the client key (current or retained previous epoch) matching the
 /// packet header's `session_key_id`, so a packet sealed under the pre-rekey key
 /// during the grace window still decrypts while an expired/stale epoch is
@@ -3841,7 +3906,7 @@ mod tests {
     use super::*;
 
     #[test]
-    fn persists_all_sessions_when_enabled() {
+    fn persists_terminal_sessions_when_enabled() {
         let (pty_tx, _rx) = mpsc::unbounded_channel();
         let config = ServerConfig {
             persist_sessions: true,
@@ -3851,7 +3916,7 @@ mod tests {
         let state = ServerState::new(config, [0u8; 32], pty_tx);
         assert!(state.should_persist_session("default")); // prewarmed
         assert!(state.should_persist_session("work")); // explicitly named
-        assert!(state.should_persist_session("term-1781470634216-76685")); // implicit reconnect
+        assert!(state.should_persist_session("term-1781470634216-76685")); // one-off update-survivable terminal
     }
 
     #[test]
@@ -4006,6 +4071,12 @@ mod tests {
             Some((key, "work".to_string()))
         );
         assert!(state.client_mut(&client_id).is_some());
+        state.sessions.get_mut("work").unwrap().empty_since = Some(Instant::now());
+        state.insert_client("work", [7u8; 16], test_client_state([9u8; 32]));
+        assert!(
+            state.sessions["work"].empty_since.is_none(),
+            "reattach must clear empty-since so cleanup cannot reap a live session"
+        );
 
         // Removing purges both the session map and the index.
         assert!(state.remove_client_everywhere(&client_id));
@@ -4045,6 +4116,83 @@ mod tests {
         assert!(locked.lookup_client(&client_id).is_none());
     }
 
+    #[test]
+    fn implicit_empty_session_timeout_is_bounded_for_update_reconnect() {
+        let configured = Duration::from_secs(2_592_000);
+        let implicit = protocol::generate_implicit_session_name();
+        assert_eq!(
+            empty_session_timeout(&implicit, configured),
+            Duration::from_secs(IMPLICIT_EMPTY_SESSION_GRACE_SECS)
+        );
+        assert_eq!(
+            empty_session_timeout("work", configured),
+            configured,
+            "named sessions keep the normal long timeout"
+        );
+        assert_eq!(
+            empty_session_timeout(&implicit, Duration::from_secs(1)),
+            Duration::from_secs(1),
+            "tests/admins can still configure a shorter timeout"
+        );
+    }
+
+    #[tokio::test]
+    async fn forged_plaintext_detach_does_not_remove_client() {
+        let (pty_tx, _pty_rx) = mpsc::unbounded_channel();
+        let mut state = ServerState::new(ServerConfig::default(), [0u8; 32], pty_tx);
+        state
+            .ensure_session("work", 80, 24, "forward-only", &[])
+            .unwrap();
+        let client_id = [21u8; 16];
+        let session_key = [22u8; 32];
+        state.insert_client("work", client_id, test_client_state(session_key));
+        let state = Arc::new(Mutex::new(state));
+        let mut packet = protocol::decode(
+            &protocol::encode_plain(PacketKind::Detach, client_id, 1, 0, b"").unwrap(),
+        )
+        .unwrap();
+        packet.header.session_key_id = protocol::session_key_id(&session_key);
+
+        assert!(handle_detach(&state, &packet).await.is_err());
+
+        let locked = state.lock().unwrap();
+        assert!(
+            locked.lookup_client(&client_id).is_some(),
+            "plaintext detach with a copied key id must not remove a live client"
+        );
+    }
+
+    #[tokio::test]
+    async fn authenticated_detach_removes_client() {
+        let (pty_tx, _pty_rx) = mpsc::unbounded_channel();
+        let mut state = ServerState::new(ServerConfig::default(), [0u8; 32], pty_tx);
+        state
+            .ensure_session("work", 80, 24, "forward-only", &[])
+            .unwrap();
+        let client_id = [23u8; 16];
+        let session_key = [24u8; 32];
+        state.insert_client("work", client_id, test_client_state(session_key));
+        let state = Arc::new(Mutex::new(state));
+        let packet = protocol::decode(
+            &protocol::encode_encrypted(
+                PacketKind::Detach,
+                client_id,
+                1,
+                0,
+                &session_key,
+                CLIENT_TO_SERVER,
+                b"",
+            )
+            .unwrap(),
+        )
+        .unwrap();
+
+        handle_detach(&state, &packet).await.unwrap();
+
+        let locked = state.lock().unwrap();
+        assert!(locked.lookup_client(&client_id).is_none());
+    }
+
     #[tokio::test]
     async fn duplicate_stream_open_for_open_stream_resends_ok() {
         let (pty_tx, _pty_rx) = mpsc::unbounded_channel();
@@ -4523,4 +4671,25 @@ mod tests {
         };
         remote_bind_allowed(&config, "0.0.0.0").unwrap();
     }
+
+    #[cfg(unix)]
+    #[test]
+    fn server_udp_send_classifier_treats_network_roaming_errors_as_transient() {
+        for code in [
+            libc::EADDRNOTAVAIL,
+            libc::ECONNREFUSED,
+            libc::ECONNRESET,
+            libc::EHOSTDOWN,
+            libc::EHOSTUNREACH,
+            libc::ENETDOWN,
+            libc::ENETRESET,
+            libc::ENETUNREACH,
+            libc::ETIMEDOUT,
+        ] {
+            assert!(
+                is_transient_udp_send_error(&std::io::Error::from_raw_os_error(code)),
+                "expected errno {code} to be transient"
+            );
+        }
+    }
 }

src/lib.rs

@@ -23,3 +23,4 @@ pub mod pty;
 pub mod server;
 pub mod ssh_agent;
 pub mod transport;
+pub mod udp;

src/persist.rs

@@ -435,6 +435,10 @@ pub fn run_holder(
         let mut cmd = [0u8; 1];
         match stream.read(&mut cmd) {
             Ok(1) if cmd[0] == HOLDER_CMD_SHUTDOWN => {
+                if shell_pid > 0 {
+                    let _ = unsafe { libc::kill(shell_pid, libc::SIGHUP) };
+                    let _ = unsafe { libc::kill(shell_pid, libc::SIGTERM) };
+                }
                 let _ = std::fs::remove_dir_all(runtime_dir);
                 std::process::exit(0);
             }

src/protocol.rs

@@ -280,7 +280,7 @@ pub fn decode(input: &[u8]) -> Result<Packet> {
 
 pub fn decrypt_body(packet: &Packet, key: &[u8; 32], direction: u32) -> Result<Vec<u8>> {
     if packet.header.flags & 1 == 0 {
-        return Ok(packet.body.clone());
+        bail!("packet body is not encrypted");
     }
     let nonce = crypto::nonce_from(direction, packet.header.seq);
     let aad = packet.header.aad();

src/server.rs

@@ -13,6 +13,7 @@ use crate::transport::{
     DoshTransport, SessionEvent, SessionRole, SessionTransportConfig, TransportConfig,
     service_name_from_target,
 };
+use crate::udp::is_transient_udp_send_error;
 use anyhow::{Context, Result, anyhow, bail};
 use ed25519_dalek::SigningKey;
 use std::collections::{HashMap, HashSet};
@@ -97,6 +98,7 @@ pub enum DoshServerEvent {
     Ignored,
 }
 
+#[derive(Debug, Clone)]
 struct PendingServerAuth {
     client: native::NativeClientHello,
     server: NativeServerHello,
@@ -255,7 +257,7 @@ impl DoshServer {
         };
         let body = protocol::to_body(&NativeServerHelloBody { hello })?;
         let out = protocol::encode_plain(PacketKind::NativeServerHello, pending_id, 1, 0, &body)?;
-        self.socket.send_to(&out, peer).await?;
+        let _ = send_udp(&self.socket, &out, peer).await?;
         Ok(())
     }
 
@@ -327,7 +329,7 @@ impl DoshServer {
         peer: SocketAddr,
         packet: &protocol::Packet,
     ) -> Result<DoshServerEvent> {
-        let Some(pending) = self.pending.remove(&packet.header.conn_id) else {
+        let Some(pending) = self.pending.get(&packet.header.conn_id).cloned() else {
             self.send_reject(peer, packet.header.conn_id, "unknown native auth")
                 .await?;
             return Ok(DoshServerEvent::Ignored);
@@ -338,13 +340,28 @@ impl DoshServer {
             return Ok(DoshServerEvent::Ignored);
         }
         if pending.created_at.elapsed() > self.config.auth_timeout {
+            self.pending.remove(&packet.header.conn_id);
             self.send_reject(peer, packet.header.conn_id, "native auth expired")
                 .await?;
             return Ok(DoshServerEvent::Ignored);
         }
 
-        let body = protocol::decrypt_body(packet, &pending.session_key, CLIENT_TO_SERVER)?;
-        let req: NativeUserAuthBody = protocol::from_body(&body)?;
+        let body = match protocol::decrypt_body(packet, &pending.session_key, CLIENT_TO_SERVER) {
+            Ok(body) => body,
+            Err(_) => {
+                self.send_reject(peer, packet.header.conn_id, "native auth decrypt failed")
+                    .await?;
+                return Ok(DoshServerEvent::Ignored);
+            }
+        };
+        let req: NativeUserAuthBody = match protocol::from_body(&body) {
+            Ok(req) => req,
+            Err(_) => {
+                self.send_reject(peer, packet.header.conn_id, "invalid native auth body")
+                    .await?;
+                return Ok(DoshServerEvent::Ignored);
+            }
+        };
         let services = match self.verify_auth_and_services(&pending, &req.auth) {
             Ok(services) => services,
             Err(err) => {
@@ -353,6 +370,7 @@ impl DoshServer {
                 return Ok(DoshServerEvent::Ignored);
             }
         };
+        self.pending.remove(&packet.header.conn_id);
 
         let conn_id = crypto::random_16();
         let key_id = protocol::session_key_id(&pending.session_key);
@@ -381,7 +399,7 @@ impl DoshServer {
             SERVER_TO_CLIENT,
             &body,
         )?;
-        self.socket.send_to(&out, peer).await?;
+        let _ = send_udp(&self.socket, &out, peer).await?;
 
         let transport = DoshTransport::new(
             Arc::clone(&self.socket),
@@ -428,7 +446,7 @@ impl DoshServer {
             reason: reason.to_string(),
         })?;
         let out = protocol::encode_plain(PacketKind::AttachReject, conn_id, 0, 0, &body)?;
-        self.socket.send_to(&out, peer).await?;
+        let _ = send_udp(&self.socket, &out, peer).await?;
         Ok(())
     }
 
@@ -439,6 +457,14 @@ impl DoshServer {
     }
 }
 
+async fn send_udp(socket: &UdpSocket, packet: &[u8], peer: SocketAddr) -> Result<bool> {
+    match socket.send_to(packet, peer).await {
+        Ok(_) => Ok(true),
+        Err(err) if is_transient_udp_send_error(&err) => Ok(false),
+        Err(err) => Err(err).with_context(|| format!("send UDP packet to {peer}")),
+    }
+}
+
 fn validate_requested_services(
     allowed_services: &HashSet<String>,
     requests: &[ForwardingRequest],
@@ -477,6 +503,7 @@ mod tests {
     use super::*;
     use crate::client::DoshClient;
     use crate::config::{ClientConfig, HostsConfig};
+    use crate::protocol::{CLIENT_TO_SERVER, NativeUserAuthBody};
     use crate::transport::TransportEvent;
     use ed25519_dalek::SigningKey;
 
@@ -588,4 +615,91 @@ mod tests {
             }
         }
     }
+
+    #[tokio::test]
+    async fn bad_native_auth_does_not_consume_pending_challenge() {
+        let dir = tempfile::tempdir().unwrap();
+        let host_key = dir.path().join("host_key");
+        let authorized_keys = dir.path().join("authorized_keys");
+        let signing = SigningKey::from_bytes(&[93u8; 32]);
+        let keypair = ssh_key::private::Ed25519Keypair::from(&signing);
+        let private =
+            ssh_key::PrivateKey::new(ssh_key::private::KeypairData::from(keypair), "").unwrap();
+        std::fs::write(
+            &authorized_keys,
+            format!("{}\n", private.public_key().to_openssh().unwrap()),
+        )
+        .unwrap();
+
+        let server_config = ServerConfig {
+            host_key: host_key.to_string_lossy().to_string(),
+            authorized_keys: vec![authorized_keys.to_string_lossy().to_string()],
+            ..ServerConfig::default()
+        };
+        let server_config = DoshServerConfig::new(server_config)
+            .bind_addr("127.0.0.1:0".parse().unwrap())
+            .require_current_user(false);
+        let mut server = DoshServer::bind(server_config).await.unwrap();
+        let peer: SocketAddr = "127.0.0.1:9".parse().unwrap();
+        let (client_secret, client_public) = native::generate_native_ephemeral();
+        let hello = native::NativeClientHello {
+            protocol_version: native::NATIVE_PROTOCOL_VERSION,
+            client_random: crypto::random_32(),
+            client_ephemeral_public: client_public,
+            requested_host: "127.0.0.1".to_string(),
+            requested_user: "sdk-user".to_string(),
+            requested_session: "test".to_string(),
+            requested_mode: "forward-only".to_string(),
+            terminal_size: (80, 24),
+            supported_aead: vec!["chacha20poly1305".to_string()],
+            supported_user_key_algorithms: vec!["ssh-ed25519".to_string()],
+            cached_host_key_fingerprint: None,
+            attach_ticket_envelope: None,
+            requested_env: Vec::new(),
+        };
+        let (pending_id, server_hello) = server.build_server_hello(hello.clone(), peer).unwrap();
+        let session_key = native::derive_native_session_key(
+            &client_secret,
+            server_hello.server_ephemeral_public,
+            &hello,
+            &server_hello,
+        )
+        .unwrap();
+        let auth = native::sign_user_auth(&signing, &hello, &server_hello, Vec::new()).unwrap();
+        let body = protocol::to_body(&NativeUserAuthBody { auth }).unwrap();
+        let bad = protocol::encode_encrypted(
+            PacketKind::NativeUserAuth,
+            pending_id,
+            2,
+            1,
+            &[7u8; 32],
+            CLIENT_TO_SERVER,
+            &body,
+        )
+        .unwrap();
+        let bad = protocol::decode(&bad).unwrap();
+
+        assert!(matches!(
+            server.handle_user_auth(peer, &bad).await.unwrap(),
+            DoshServerEvent::Ignored
+        ));
+        assert!(server.pending.contains_key(&pending_id));
+
+        let valid = protocol::encode_encrypted(
+            PacketKind::NativeUserAuth,
+            pending_id,
+            2,
+            1,
+            &session_key,
+            CLIENT_TO_SERVER,
+            &body,
+        )
+        .unwrap();
+        let valid = protocol::decode(&valid).unwrap();
+        assert!(matches!(
+            server.handle_user_auth(peer, &valid).await.unwrap(),
+            DoshServerEvent::Accepted(_)
+        ));
+        assert!(!server.pending.contains_key(&pending_id));
+    }
 }

src/transport.rs

@@ -19,6 +19,7 @@ use crate::protocol::{
     self, CLIENT_TO_SERVER, PacketKind, ReplayWindow, SERVER_TO_CLIENT, StreamClose, StreamData,
     StreamOpen, StreamOpenOk, StreamOpenReject, StreamWindowAdjust,
 };
+use crate::udp::is_transient_udp_send_error;
 use anyhow::{Result, anyhow, bail};
 use std::collections::{BTreeMap, HashMap, HashSet, VecDeque};
 use std::net::SocketAddr;
@@ -348,8 +349,8 @@ impl StreamMux {
         &mut self,
         adjust: StreamWindowAdjust,
     ) -> Result<Vec<OutgoingStreamPacket>> {
-        self.ack_data(adjust.stream_id, adjust.received_offset);
-        self.add_credit(adjust.stream_id, adjust.bytes as usize);
+        let acked_bytes = self.ack_data(adjust.stream_id, adjust.received_offset);
+        self.add_credit(adjust.stream_id, acked_bytes);
         self.flush_pending_data(adjust.stream_id)
     }
 
@@ -556,9 +557,9 @@ impl StreamMux {
         (chunks, consumed, *expected)
     }
 
-    fn ack_data(&mut self, stream_id: u64, received_offset: u64) {
+    fn ack_data(&mut self, stream_id: u64, received_offset: u64) -> usize {
         let Some(sent) = self.sent_data.get_mut(&stream_id) else {
-            return;
+            return 0;
         };
         let acked_offsets = sent
             .iter()
@@ -567,12 +568,16 @@ impl StreamMux {
                 (end <= received_offset).then_some(*offset)
             })
             .collect::<Vec<_>>();
+        let mut acked_bytes = 0usize;
         for offset in acked_offsets {
-            sent.remove(&offset);
+            if let Some(chunk) = sent.remove(&offset) {
+                acked_bytes = acked_bytes.saturating_add(chunk.bytes.len());
+            }
         }
         if sent.is_empty() {
             self.sent_data.remove(&stream_id);
         }
+        acked_bytes
     }
 
     fn add_credit(&mut self, stream_id: u64, bytes: usize) {
@@ -732,11 +737,17 @@ impl DoshTransport {
             if packet.header.conn_id != self.conn_id {
                 continue;
             }
+            let plain = match protocol::decrypt_body(
+                &packet,
+                &self.session_key,
+                self.role.recv_direction(),
+            ) {
+                Ok(plain) => plain,
+                Err(_) => continue,
+            };
             if !self.replay.accept(packet.header.seq) {
                 continue;
             }
-            let plain =
-                protocol::decrypt_body(&packet, &self.session_key, self.role.recv_direction())?;
             self.peer_addr = peer;
             self.ack_seq = self.ack_seq.max(packet.header.seq);
             self.last_contact = Instant::now();
@@ -751,14 +762,21 @@ impl DoshTransport {
         datagram: &[u8],
         peer: SocketAddr,
     ) -> Result<SessionEvent> {
-        let packet = protocol::decode(datagram)?;
+        let packet = match protocol::decode(datagram) {
+            Ok(packet) => packet,
+            Err(_) => return Ok(SessionEvent::Ignored),
+        };
         if packet.header.conn_id != self.conn_id {
             return Ok(SessionEvent::Ignored);
         }
+        let plain =
+            match protocol::decrypt_body(&packet, &self.session_key, self.role.recv_direction()) {
+                Ok(plain) => plain,
+                Err(_) => return Ok(SessionEvent::Ignored),
+            };
         if !self.replay.accept(packet.header.seq) {
             return Ok(SessionEvent::Ignored);
         }
-        let plain = protocol::decrypt_body(&packet, &self.session_key, self.role.recv_direction())?;
         self.peer_addr = peer;
         self.ack_seq = self.ack_seq.max(packet.header.seq);
         self.last_contact = Instant::now();
@@ -815,7 +833,12 @@ impl DoshTransport {
 
     async fn send_kind(&mut self, kind: PacketKind, body: &[u8]) -> Result<()> {
         let encoded = self.encode_kind(kind, body)?;
-        self.socket.send_to(&encoded, self.peer_addr).await?;
+        if let Err(err) = self.socket.send_to(&encoded, self.peer_addr).await {
+            if is_transient_udp_send_error(&err) {
+                return Ok(());
+            }
+            return Err(err.into());
+        }
         Ok(())
     }
 
@@ -980,6 +1003,30 @@ mod tests {
         assert_eq!(data.bytes, b"!");
     }
 
+    #[test]
+    fn cumulative_window_adjust_restores_credit_even_if_delta_was_lost() {
+        let mut mux = StreamMux::new(TransportConfig {
+            initial_window: 5,
+            retransmit_after: Duration::from_secs(1),
+            ..TransportConfig::default()
+        });
+        mux.open_stream(1, "@dosh-test", 0).unwrap();
+        mux.handle_open_ok(StreamOpenOk { stream_id: 1 }).unwrap();
+        assert_eq!(mux.send_data(1, b"hello".to_vec()).unwrap().len(), 1);
+        assert_eq!(mux.send_credit[&1], 0);
+
+        let flushed = mux
+            .handle_window_adjust(StreamWindowAdjust {
+                stream_id: 1,
+                received_offset: 5,
+                bytes: 0,
+            })
+            .unwrap();
+        assert!(flushed.is_empty());
+        assert_eq!(mux.send_credit[&1], 5);
+        assert_eq!(mux.send_data(1, b"!".to_vec()).unwrap().len(), 1);
+    }
+
     #[test]
     fn handle_packet_decodes_and_dispatches_stream_packets() {
         let mut mux = StreamMux::new(TransportConfig::default());
@@ -1096,4 +1143,61 @@ mod tests {
         assert!(matches!(server.recv().await.unwrap(), SessionEvent::Ping));
         assert_eq!(server.peer_addr(), roaming_addr);
     }
+
+    #[tokio::test]
+    async fn unauthenticated_datagram_does_not_poison_replay_window() {
+        let socket = UdpSocket::bind("127.0.0.1:0").await.unwrap();
+        let peer: SocketAddr = "127.0.0.1:9".parse().unwrap();
+        let key = [11u8; 32];
+        let wrong_key = [12u8; 32];
+        let conn_id = [4u8; 16];
+        let mut transport = DoshTransport::new_owned(
+            socket,
+            SessionTransportConfig {
+                role: SessionRole::Client,
+                conn_id,
+                session_key: key,
+                peer_addr: peer,
+                initial_send_seq: 1,
+                initial_ack: 0,
+                stream: TransportConfig::default(),
+            },
+        );
+        let bad = protocol::encode_encrypted(
+            PacketKind::Pong,
+            conn_id,
+            9,
+            0,
+            &wrong_key,
+            SERVER_TO_CLIENT,
+            b"",
+        )
+        .unwrap();
+        let valid = protocol::encode_encrypted(
+            PacketKind::Pong,
+            conn_id,
+            9,
+            0,
+            &key,
+            SERVER_TO_CLIENT,
+            b"",
+        )
+        .unwrap();
+
+        assert!(matches!(
+            transport
+                .handle_datagram(b"not a dosh packet", peer)
+                .await
+                .unwrap(),
+            SessionEvent::Ignored
+        ));
+        assert!(matches!(
+            transport.handle_datagram(&bad, peer).await.unwrap(),
+            SessionEvent::Ignored
+        ));
+        assert!(matches!(
+            transport.handle_datagram(&valid, peer).await.unwrap(),
+            SessionEvent::Pong
+        ));
+    }
 }

src/udp.rs

@@ -0,0 +1,98 @@
+//! UDP error classification shared by Dosh clients, servers, and embedders.
+
+pub fn is_transient_udp_send_error(err: &std::io::Error) -> bool {
+    matches!(
+        err.kind(),
+        std::io::ErrorKind::Interrupted
+            | std::io::ErrorKind::TimedOut
+            | std::io::ErrorKind::WouldBlock
+    ) || err.raw_os_error().is_some_and(is_transient_udp_os_error)
+}
+
+#[cfg(unix)]
+fn is_transient_udp_os_error(code: i32) -> bool {
+    matches!(
+        code,
+        libc::EADDRNOTAVAIL
+            | libc::ECONNREFUSED
+            | libc::ECONNRESET
+            | libc::EHOSTDOWN
+            | libc::EHOSTUNREACH
+            | libc::ENETDOWN
+            | libc::ENETRESET
+            | libc::ENETUNREACH
+            | libc::ETIMEDOUT
+    )
+}
+
+#[cfg(windows)]
+fn is_transient_udp_os_error(code: i32) -> bool {
+    matches!(
+        code,
+        10049  // WSAEADDRNOTAVAIL
+            | 10050 // WSAENETDOWN
+            | 10051 // WSAENETUNREACH
+            | 10052 // WSAENETRESET
+            | 10054 // WSAECONNRESET
+            | 10060 // WSAETIMEDOUT
+            | 10061 // WSAECONNREFUSED
+            | 10064 // WSAEHOSTDOWN
+            | 10065 // WSAEHOSTUNREACH
+    )
+}
+
+#[cfg(not(any(unix, windows)))]
+fn is_transient_udp_os_error(_code: i32) -> bool {
+    false
+}
+
+#[cfg(test)]
+mod tests {
+    use super::is_transient_udp_send_error;
+
+    #[test]
+    fn classifies_portable_transient_send_errors() {
+        for kind in [
+            std::io::ErrorKind::Interrupted,
+            std::io::ErrorKind::TimedOut,
+            std::io::ErrorKind::WouldBlock,
+        ] {
+            assert!(is_transient_udp_send_error(&std::io::Error::from(kind)));
+        }
+        assert!(!is_transient_udp_send_error(&std::io::Error::from(
+            std::io::ErrorKind::PermissionDenied,
+        )));
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn classifies_unix_network_churn_as_transient() {
+        for code in [
+            libc::EADDRNOTAVAIL,
+            libc::ECONNREFUSED,
+            libc::ECONNRESET,
+            libc::EHOSTDOWN,
+            libc::EHOSTUNREACH,
+            libc::ENETDOWN,
+            libc::ENETRESET,
+            libc::ENETUNREACH,
+            libc::ETIMEDOUT,
+        ] {
+            assert!(is_transient_udp_send_error(
+                &std::io::Error::from_raw_os_error(code)
+            ));
+        }
+    }
+
+    #[cfg(windows)]
+    #[test]
+    fn classifies_windows_network_churn_as_transient() {
+        for code in [
+            10049, 10050, 10051, 10052, 10054, 10060, 10061, 10064, 10065,
+        ] {
+            assert!(is_transient_udp_send_error(
+                &std::io::Error::from_raw_os_error(code)
+            ));
+        }
+    }
+}

tests/hostile_network.rs

@@ -258,7 +258,8 @@ fn relay_loop(
     let mut upstream = new_upstream();
     let mut client_addr: Option<SocketAddr> = None;
     let mut rng = StdRng::seed_from_u64(seed);
-    let mut held: Option<(Vec<u8>, bool)> = None; // (packet, is_c2s) held for reorder
+    let mut held_c2s: Option<Vec<u8>> = None;
+    let mut held_s2c: Option<Vec<u8>> = None;
     let mut buf = [0u8; 65535];
 
     loop {
@@ -288,10 +289,9 @@ fn relay_loop(
                         &upstream,
                         server_addr,
                         packet,
-                        true,
                         &controls,
                         &mut rng,
-                        &mut held,
+                        &mut held_c2s,
                     );
                 }
             }
@@ -310,7 +310,12 @@ fn relay_loop(
                     let drop_pct = controls.drop_s2c_percent.load(Ordering::SeqCst);
                     if drop_pct == 0 || rng.gen_range(0..100) >= drop_pct {
                         forward_with_effects(
-                            &front, dst, packet, false, &controls, &mut rng, &mut held,
+                            &front,
+                            dst,
+                            packet,
+                            &controls,
+                            &mut rng,
+                            &mut held_s2c,
                         );
                     }
                 }
@@ -336,23 +341,22 @@ fn forward_with_effects(
     out: &UdpSocket,
     dst: SocketAddr,
     packet: Vec<u8>,
-    is_c2s: bool,
     controls: &RelayControls,
     rng: &mut StdRng,
-    held: &mut Option<(Vec<u8>, bool)>,
+    held: &mut Option<Vec<u8>>,
 ) {
     // Reorder: if armed, hold this packet and release the previously held one
     // afterward (so two consecutive packets swap order).
     if controls.reorder_next.swap(false, Ordering::SeqCst) {
-        if let Some((prev, _)) = held.take() {
+        if let Some(prev) = held.take() {
             let _ = out.send_to(&packet, dst);
             let _ = out.send_to(&prev, dst);
             return;
         }
-        *held = Some((packet, is_c2s));
+        *held = Some(packet);
         return;
     }
-    if let Some((prev, _)) = held.take() {
+    if let Some(prev) = held.take() {
         let _ = out.send_to(&prev, dst);
     }
 
@@ -767,6 +771,46 @@ fn wait_for_text(socket: &UdpSocket, key: &[u8; 32], needle: &str, millis: u64)
     acc.contains(needle)
 }
 
+fn wait_for_text_with_ack(
+    socket: &UdpSocket,
+    relay: &Relay,
+    client_id: [u8; 16],
+    seq: &mut u64,
+    key: &[u8; 32],
+    needle: &str,
+    millis: u64,
+) -> bool {
+    let prev = socket.read_timeout().unwrap();
+    socket
+        .set_read_timeout(Some(Duration::from_millis(100)))
+        .unwrap();
+    let deadline = Instant::now() + Duration::from_millis(millis);
+    let mut acc = String::new();
+    while Instant::now() < deadline {
+        if let Some((_header, frame)) = recv_frame(socket, key) {
+            acc.push_str(&String::from_utf8_lossy(&frame.bytes));
+            let ack = protocol::encode_encrypted(
+                PacketKind::Ack,
+                client_id,
+                *seq,
+                frame.output_seq,
+                key,
+                CLIENT_TO_SERVER,
+                b"",
+            )
+            .unwrap();
+            *seq += 1;
+            socket.send_to(&ack, relay.front_addr()).unwrap();
+            if acc.contains(needle) {
+                socket.set_read_timeout(prev).unwrap();
+                return true;
+            }
+        }
+    }
+    socket.set_read_timeout(prev).unwrap();
+    acc.contains(needle)
+}
+
 #[test]
 fn session_survives_packet_loss_and_reorder() {
     let dir = tempfile::tempdir().unwrap();
@@ -820,6 +864,94 @@ fn session_survives_packet_loss_and_reorder() {
     );
 }
 
+#[test]
+#[ignore = "30-minute hostile-network TUI soak; run with `DOSH_BADNET_SOAK_SECONDS=1800 cargo test --test hostile_network bad_network_tui_work_soak_30m -- --ignored --nocapture`"]
+fn bad_network_tui_work_soak_30m() {
+    let soak_secs = std::env::var("DOSH_BADNET_SOAK_SECONDS")
+        .ok()
+        .and_then(|value| value.parse::<u64>().ok())
+        .unwrap_or(30 * 60);
+    let dir = tempfile::tempdir().unwrap();
+    let port = free_udp_port();
+    let config = write_server_config(&dir, port);
+    let mut server = start_server(&dir, &config);
+    let relay = Relay::spawn(port, 0xBADC0DEu64);
+    let (socket, bootstrap, ok) = attach_through_relay(&config, &relay);
+
+    relay.set_drop_c2s(15);
+    relay.set_drop_s2c(20);
+    relay.set_dup(10);
+
+    let deadline = Instant::now() + Duration::from_secs(soak_secs);
+    let mut seq = 2u64;
+    let mut iteration = 0u64;
+    while Instant::now() < deadline {
+        if iteration.is_multiple_of(2) {
+            relay.arm_reorder();
+        }
+        if iteration > 0 && iteration.is_multiple_of(7) {
+            let _ = relay.rebind_upstream();
+        }
+        if iteration > 0 && iteration.is_multiple_of(11) {
+            relay.set_drop_s2c(100);
+            thread::sleep(Duration::from_millis(750));
+            relay.set_drop_s2c(20);
+        }
+
+        let marker = format!("DOSH_BADNET_TUI_{iteration}");
+        let command = format!(
+            "stty -echo; \
+             printf '\\033[?1049h\\033[?2026h\\033[?25l'; \
+             for i in 1 2 3 4 5 6; do \
+               printf '\\033[%s;4H{} ⠀⠁⠃⠇⡇⣇⣧⣷⣿ █▇▆▅▄▃▂▁ %s\\033[0m' \"$i\" \"$i\"; \
+             done; \
+             printf '\\033[?25h\\033[?2026l\\033[?1049l'\n",
+            marker
+        );
+
+        let step_deadline = Instant::now() + Duration::from_secs(8);
+        let mut seen = false;
+        let mut attempts = 0;
+        while Instant::now() < step_deadline && attempts < 12 {
+            send_input(
+                &socket,
+                &relay,
+                ok.client_id,
+                seq,
+                0,
+                &bootstrap.session_key,
+                command.as_bytes(),
+            );
+            seq += 1;
+            attempts += 1;
+            if wait_for_text_with_ack(
+                &socket,
+                &relay,
+                ok.client_id,
+                &mut seq,
+                &bootstrap.session_key,
+                &marker,
+                500,
+            ) {
+                seen = true;
+                break;
+            }
+        }
+        assert!(
+            seen,
+            "TUI marker {marker} did not arrive during hostile-network soak"
+        );
+
+        iteration += 1;
+        thread::sleep(Duration::from_millis(500));
+    }
+
+    relay.clear_impairments();
+    drop(relay);
+    let _ = server.kill();
+    let _ = server.wait();
+}
+
 #[test]
 fn duplicated_and_replayed_input_is_applied_at_most_once() {
     let dir = tempfile::tempdir().unwrap();

tests/integration_smoke.rs

@@ -18,7 +18,8 @@ use dosh::crypto;
 use dosh::native::{host_public_key, load_or_create_host_key, ssh_ed25519_public_blob, trust_host};
 use dosh::protocol::{
     self, AttachOk, AttachReject, BootstrapAttachRequest, CLIENT_TO_SERVER, Frame, Input,
-    PacketKind, Resize, ResumeRequest, SERVER_TO_CLIENT,
+    PacketKind, Resize, ResumeRequest, SERVER_TO_CLIENT, TicketAttachBody, TicketAttachEnvelope,
+    TicketAttachOkEnvelope,
 };
 use ed25519_dalek::{Signer, SigningKey, VerifyingKey};
 
@@ -530,6 +531,75 @@ fn direct_attach_session(
     (socket, bootstrap, ok)
 }
 
+fn ticket_attach_session(
+    socket: &UdpSocket,
+    port: u16,
+    bootstrap: &dosh::auth::BootstrapResponse,
+    session: &str,
+    mode: &str,
+) -> AttachOk {
+    let client_nonce = crypto::random_12();
+    let request_key = crypto::hkdf32(
+        &bootstrap.attach_ticket_psk,
+        &client_nonce,
+        b"dosh/ticket-attach-request/v1",
+    )
+    .unwrap();
+    let body = protocol::to_body(&TicketAttachBody {
+        session: session.to_string(),
+        mode: mode.to_string(),
+        cols: 80,
+        rows: 24,
+        requested_env: Vec::new(),
+    })
+    .unwrap();
+    let ciphertext = crypto::seal(
+        &request_key,
+        &client_nonce,
+        b"dosh-ticket-attach-request-v1",
+        &body,
+    )
+    .unwrap();
+    let envelope = TicketAttachEnvelope {
+        ticket: bootstrap.attach_ticket.clone(),
+        client_nonce,
+        ciphertext,
+    };
+    let packet = protocol::encode_plain(
+        PacketKind::TicketAttachRequest,
+        [0u8; 16],
+        1,
+        0,
+        &protocol::to_body(&envelope).unwrap(),
+    )
+    .unwrap();
+    socket
+        .send_to(&packet, format!("127.0.0.1:{port}"))
+        .unwrap();
+    let mut buf = [0u8; 65535];
+    let (n, _) = socket.recv_from(&mut buf).unwrap();
+    let packet = protocol::decode(&buf[..n]).unwrap();
+    assert_eq!(packet.header.kind, PacketKind::AttachOk);
+    let envelope: TicketAttachOkEnvelope = protocol::from_body(&packet.body).unwrap();
+    let mut salt = Vec::with_capacity(24);
+    salt.extend_from_slice(&client_nonce);
+    salt.extend_from_slice(&envelope.server_nonce);
+    let response_key = crypto::hkdf32(
+        &bootstrap.attach_ticket_psk,
+        &salt,
+        b"dosh/ticket-attach-ok/v1",
+    )
+    .unwrap();
+    let plain = crypto::open(
+        &response_key,
+        &envelope.server_nonce,
+        b"dosh-ticket-attach-ok-v1",
+        &envelope.ciphertext,
+    )
+    .unwrap();
+    protocol::from_body(&plain).unwrap()
+}
+
 #[allow(clippy::too_many_arguments)]
 fn send_encrypted(
     socket: &UdpSocket,
@@ -1392,18 +1462,10 @@ fn server_retransmits_unacked_output_frames() {
     );
 
     let mut seen = Vec::new();
-    let mut duplicate = None;
-    let deadline = std::time::Instant::now() + Duration::from_secs(3);
+    let deadline = std::time::Instant::now() + Duration::from_secs(4);
     while std::time::Instant::now() < deadline {
         if let Some((header, frame)) = recv_frame(&socket, &bootstrap.session_key) {
             if String::from_utf8_lossy(&frame.bytes).contains("DOSH_RETRANSMIT") {
-                if seen
-                    .iter()
-                    .any(|(_, output_seq)| *output_seq == frame.output_seq)
-                {
-                    duplicate = Some((header.seq, frame.output_seq));
-                    break;
-                }
                 seen.push((header.seq, frame.output_seq));
             }
         }
@@ -1412,9 +1474,18 @@ fn server_retransmits_unacked_output_frames() {
     let _ = server.kill();
     let _ = server.wait();
 
+    let max_same_output_seq = seen
+        .iter()
+        .map(|(_, output_seq)| {
+            seen.iter()
+                .filter(|(_, candidate)| candidate == output_seq)
+                .count()
+        })
+        .max()
+        .unwrap_or(0);
     assert!(
-        duplicate.is_some(),
-        "expected retransmitted same output seq, seen={seen:?}"
+        max_same_output_seq >= 10,
+        "expected retransmission to continue past the old 8-attempt ceiling, seen={seen:?}"
     );
 }
 
@@ -2518,6 +2589,9 @@ fn write_persistent_server_config(dir: &tempfile::TempDir, port: u16) -> std::pa
     let mut raw = fs::read_to_string(&config).unwrap();
     // The base writer hardcodes `persist_sessions = false`; flip it on.
     raw = raw.replace("persist_sessions = false", "persist_sessions = true");
+    // Persistence tests attach the sessions they need explicitly. Leaving the
+    // base prewarm on here creates a detached default holder in every test.
+    raw = raw.replace("prewarm_sessions = [\"default\"]", "prewarm_sessions = []");
     fs::write(&config, raw).unwrap();
     config
 }
@@ -2553,6 +2627,45 @@ fn kill_holder(sessions_dir: &Path, session: &str) {
         .status();
 }
 
+/// Create a holder the way older Dosh builds did for implicit sessions, so
+/// startup migration can be tested without depending on the new spawn policy.
+fn spawn_legacy_holder(sessions_dir: &Path, session: &str) {
+    let server = env!("CARGO_BIN_EXE_dosh-server");
+    let runtime_dir = dosh::persist::ensure_runtime_dir(sessions_dir, session).unwrap();
+    let mut launcher = Command::new(server)
+        .arg("hold")
+        .arg("--runtime-dir")
+        .arg(&runtime_dir)
+        .arg("--session")
+        .arg(session)
+        .arg("--shell")
+        .arg("/bin/sh")
+        .arg("--cols")
+        .arg("80")
+        .arg("--rows")
+        .arg("24")
+        .stdout(Stdio::null())
+        .stderr(Stdio::null())
+        .spawn()
+        .unwrap();
+    let deadline = std::time::Instant::now() + Duration::from_secs(5);
+    while std::time::Instant::now() < deadline {
+        if holder_shell_pid(sessions_dir, session).is_some()
+            && runtime_dir.join("holder.sock").exists()
+        {
+            let _ = launcher.wait();
+            return;
+        }
+        if let Some(status) = launcher.try_wait().unwrap() {
+            panic!("legacy holder launcher exited before ready: {status}");
+        }
+        thread::sleep(Duration::from_millis(20));
+    }
+    let _ = launcher.kill();
+    let _ = launcher.wait();
+    panic!("legacy holder did not become ready");
+}
+
 /// Send one encrypted Input line and give the shell a moment to act.
 fn type_line(socket: &UdpSocket, port: u16, ok: &AttachOk, key: &[u8; 32], seq: u64, line: &str) {
     let input = Input {
@@ -2690,7 +2803,7 @@ fn session_survives_server_restart_same_shell_and_screen() {
 }
 
 #[test]
-fn implicit_session_survives_server_restart_for_same_ticket_holder() {
+fn implicit_session_survives_update_restart_via_ticket_reattach() {
     let dir = tempfile::tempdir().unwrap();
     let sessions_dir = dir.path().join("sessions");
     let port = free_udp_port();
@@ -2700,59 +2813,140 @@ fn implicit_session_survives_server_restart_for_same_ticket_holder() {
     let mut server = start_server(&dir, &config);
     let (socket, bootstrap, ok) = direct_attach_session(&config, port, "read-write", &session);
     let key = bootstrap.session_key;
+    type_line(&socket, port, &ok, &key, 2, "cd /tmp\n");
     type_line(
         &socket,
         port,
         &ok,
         &key,
-        2,
+        3,
         "export IMPLICIT_MARK=implicit_restart_42\n",
     );
-    let _ = collect_frame_text(&socket, &key, 1000);
+    type_line(
+        &socket,
+        port,
+        &ok,
+        &key,
+        4,
+        "printf 'IMPLICIT_SCREEN_MARKER\\n'\n",
+    );
+    let pre = collect_frame_text(&socket, &key, 1500);
+    assert!(
+        pre.contains("IMPLICIT_SCREEN_MARKER"),
+        "implicit marker missing before restart: {pre:?}"
+    );
     thread::sleep(Duration::from_secs(3));
 
     let shell_pid_before = holder_shell_pid(&sessions_dir, &session);
     assert!(
         shell_pid_before.is_some(),
-        "implicit sessions should get a persistent holder when persistence is enabled"
+        "implicit sessions must get temporary holders so active clients survive updates"
     );
 
     let _ = server.kill();
     let _ = server.wait();
     thread::sleep(Duration::from_millis(300));
 
-    let pid = shell_pid_before.unwrap();
-    assert!(
-        unsafe { libc::kill(pid, 0) } == 0,
-        "implicit session shell pid {pid} must survive server restart"
-    );
-
     let mut server = start_server(&dir, &config);
-    let (socket2, bootstrap2, ok2) = direct_attach_session(&config, port, "read-write", &session);
-    let key2 = bootstrap2.session_key;
+
+    let resume = ResumeRequest {
+        session: session.clone(),
+        last_rendered_seq: ok.initial_seq,
+        cols: 80,
+        rows: 24,
+    };
+    send_encrypted(
+        &socket,
+        port,
+        PacketKind::ResumeRequest,
+        ok.client_id,
+        5,
+        0,
+        &key,
+        &protocol::to_body(&resume).unwrap(),
+    );
+    let mut buf = [0u8; 65535];
+    let deadline = std::time::Instant::now() + Duration::from_secs(2);
+    loop {
+        assert!(
+            std::time::Instant::now() < deadline,
+            "old live resume did not receive unknown-client reject"
+        );
+        let (n, _) = socket.recv_from(&mut buf).unwrap();
+        let packet = protocol::decode(&buf[..n]).unwrap();
+        if packet.header.kind == PacketKind::Frame {
+            continue;
+        }
+        assert_eq!(packet.header.kind, PacketKind::AttachReject);
+        assert_eq!(packet.header.conn_id, ok.client_id);
+        break;
+    }
+
+    let ok2 = ticket_attach_session(&socket, port, &bootstrap, &session, "read-write");
+    let key2 = ok2.session_key;
+    let snapshot = String::from_utf8_lossy(&ok2.snapshot).to_string();
     type_line(
-        &socket2,
+        &socket,
         port,
         &ok2,
         &key2,
         2,
-        "printf 'IMPLICIT_MARK=%s\\n' \"$IMPLICIT_MARK\"\n",
+        "printf 'IMPLICIT_MARK=%s PWD=%s\\n' \"$IMPLICIT_MARK\" \"$PWD\"\n",
     );
-    let post = collect_frame_text(&socket2, &key2, 2000);
+    let post = collect_frame_text(&socket, &key2, 2000);
     let shell_pid_after = holder_shell_pid(&sessions_dir, &session);
 
     let _ = server.kill();
     let _ = server.wait();
     kill_holder(&sessions_dir, &session);
 
+    assert!(
+        snapshot.contains("IMPLICIT_SCREEN_MARKER"),
+        "ticket reattach snapshot must repaint the exact pre-update screen, got {snapshot:?}"
+    );
     assert_eq!(
         shell_pid_after, shell_pid_before,
-        "implicit session should re-adopt the same shell pid"
+        "implicit update reconnect must use the same holder shell"
     );
     assert!(
         post.contains("IMPLICIT_MARK=implicit_restart_42"),
-        "implicit session state must survive restart, got {post:?}"
+        "implicit session env must survive update reconnect, got {post:?}"
     );
+    assert!(
+        post.contains("PWD=/tmp"),
+        "implicit session cwd must survive update reconnect, got {post:?}"
+    );
+}
+
+#[test]
+fn startup_readopts_implicit_holders_for_update_reconnect() {
+    let dir = tempfile::tempdir().unwrap();
+    let sessions_dir = dir.path().join("sessions");
+    let port = free_udp_port();
+    let config = write_persistent_server_config(&dir, port);
+    let session = protocol::generate_implicit_session_name();
+
+    spawn_legacy_holder(&sessions_dir, &session);
+    let shell_pid = holder_shell_pid(&sessions_dir, &session).expect("legacy holder shell pid");
+    assert!(
+        unsafe { libc::kill(shell_pid, 0) } == 0,
+        "legacy holder shell must be live before startup"
+    );
+
+    let mut server = start_server(&dir, &config);
+    thread::sleep(Duration::from_millis(300));
+
+    let _ = server.kill();
+    let _ = server.wait();
+    assert!(
+        holder_shell_pid(&sessions_dir, &session).is_some(),
+        "startup must keep live implicit holders so active update reconnects can reattach"
+    );
+    assert!(
+        unsafe { libc::kill(shell_pid, 0) } == 0,
+        "startup must not shut down a live implicit holder before reconnect grace"
+    );
+    kill_holder(&sessions_dir, &session);
 }
 
 #[test]

tests/protocol_auth.rs

@@ -58,6 +58,19 @@ fn encrypted_packet_round_trips() {
     assert_eq!(plain, b"hello");
 }
 
+#[test]
+fn plaintext_packet_never_decrypts_as_authenticated_body() {
+    let key = crypto::random_32();
+    let mut decoded = protocol::decode(
+        &protocol::encode_plain(PacketKind::Input, crypto::random_16(), 1, 0, b"hello").unwrap(),
+    )
+    .unwrap();
+    decoded.header.session_key_id = protocol::session_key_id(&key);
+
+    let err = protocol::decrypt_body(&decoded, &key, CLIENT_TO_SERVER).unwrap_err();
+    assert!(err.to_string().contains("not encrypted"));
+}
+
 #[test]
 fn encrypted_packet_rejects_wrong_session_key_id_before_decrypt() {
     let key = crypto::random_32();

tests/release_scripts.rs

@@ -18,3 +18,35 @@ fn release_scripts_skip_stale_artifacts_when_auto_discovering() {
     assert!(publish.contains("skipping stale artifact"));
     assert!(publish.contains("actual=\"$(artifact_version \"$artifact\" || true)\""));
 }
+
+#[test]
+fn systemd_service_does_not_sandbox_remote_shells() {
+    let service = include_str!("../packaging/systemd/dosh-server.service");
+    let install = include_str!("../install.sh");
+    for raw in [service, install] {
+        assert!(raw.contains("KillMode=process"));
+        assert!(
+            raw.contains("Restart=always"),
+            "dosh-server should come back after accidental SIGTERM while preserving child shells"
+        );
+        for directive in [
+            "NoNewPrivileges=",
+            "PrivateTmp=",
+            "ProtectSystem=",
+            "ProtectHome=",
+            "RestrictAddressFamilies=",
+            "RestrictRealtime=",
+            "LockPersonality=",
+            "MemoryDenyWriteExecute=",
+        ] {
+            assert!(
+                !raw.contains(directive),
+                "dosh sessions are user shells; systemd sandbox directive {directive} changes terminal/app behavior"
+            );
+        }
+        assert!(
+            !raw.contains("ReadWritePaths="),
+            "remote shells must not be restricted to dosh config/data paths"
+        );
+    }
+}