feat(client): implement automatic session reconnection and disable statistics

crates/irtt-app/src/cmd/client/args.rs

@@ -55,7 +55,7 @@ pub struct ClientArgs {
         default_value = "10s",
         value_parser = parse_test_duration,
         help = "Test duration (e.g. 2.5s, 1m30s, 1h); use --continuous for continuous mode (0 is also accepted)",
-        long_help = "Test duration (e.g. 2.5s, 1m30s, 500us, 100ns, 1h); use --continuous for continuous mode (0 is also accepted).\n\nFinite runs retain exact statistics for final summaries. Continuous mode uses bounded-memory running statistics and prints a final summary when interrupted; --targets-stdin also prints the final retained summary on EOF."
+        long_help = "Test duration (e.g. 2.5s, 1m30s, 500us, 100ns, 1h); use --continuous for continuous mode (0 is also accepted).\n\nUnless --no-stats is selected, finite runs retain exact statistics for final summaries. Continuous mode uses bounded-memory running statistics and prints a final summary when interrupted; --targets-stdin also prints the final retained summary on EOF."
     )]
     pub duration: Duration,
 
@@ -63,6 +63,22 @@ pub struct ClientArgs {
     #[arg(long, conflicts_with = "duration")]
     pub continuous: bool,
 
+    /// Disable statistics collection and final summaries; retain measurement event output.
+    #[arg(long)]
+    pub no_stats: bool,
+
+    /// Retry failed, closed, or silent sessions (requires continuous mode).
+    #[arg(
+        long,
+        conflicts_with = "no_test",
+        long_help = "Retry failed initial opens and closed, failed, or silent sessions independently per target. Requires --continuous or --duration 0. Unsuccessful connections back off from 250ms to 10s. Use --reconnect-timeout to configure silence detection and --no-stats when only measurement events are needed."
+    )]
+    pub reconnect: bool,
+
+    /// Reconnect after this long without valid replies while sending probes.
+    #[arg(long, default_value = "10s", value_parser = crate::shared::client::parse_duration, requires = "reconnect")]
+    pub reconnect_timeout: Duration,
+
     #[command(flatten)]
     pub common: CommonClientArgs,
 
@@ -112,6 +128,9 @@ impl ClientArgs {
     /// Targets are only required here. `--list-columns` returns before this is
     /// called, so listing columns still needs no target.
     pub async fn prepare(&self) -> Result<ManagedRunSetup, String> {
+        if self.reconnect && !self.is_continuous() {
+            return Err("--reconnect requires --continuous or --duration 0".to_owned());
+        }
         if self.targets_stdin && !self.is_continuous() {
             return Err("--targets-stdin requires --continuous or --duration 0".to_owned());
         }

crates/irtt-app/src/cmd/client/run.rs

@@ -58,8 +58,11 @@ pub async fn run_stream(
         args.verbose,
     )
     .map_err(|err| io::Error::new(io::ErrorKind::InvalidInput, err))?;
+    if let Some(warning) = reconnect_stats_warning(&args) {
+        eprintln!("{warning}");
+    }
     if setup.stdin_controlled {
-        return run_stdin_stream(setup, output_config, shutdown).await;
+        return run_stdin_stream(setup, output_config, shutdown, &args).await;
     }
     let continuous = args.is_continuous();
     let target_count = setup.target_count();
@@ -70,12 +73,21 @@ pub async fn run_stream(
     if *shutdown.borrow() {
         return Ok(());
     }
-    let (task, handle) = ManagedClient::task(setup.managed_config(), setup.managed_targets())?;
+    let (task, handle) = ManagedClient::task(
+        {
+            let mut config = setup.managed_config();
+            config.reconnect_timeout = args.reconnect.then_some(args.reconnect_timeout);
+            config
+        },
+        setup.managed_targets(),
+    )?;
     let mut events = handle.subscribe()?;
     let mut status = handle.subscribe_status();
     let owner = ManagedWorker::start(task, handle.clone())?;
     let mut stdout = io::LineWriter::new(io::stdout().lock());
     let mut stream_output = StreamOutput {
+        continuous: args.is_continuous(),
+        collect_stats: !args.no_stats,
         config: output_config,
         header_printed: false,
         print_final_summary: false,
@@ -86,6 +98,7 @@ pub async fn run_stream(
     let mut stats = setup
         .targets
         .iter()
+        .filter(|_| !args.no_stats)
         .map(|target| {
             (
                 target.label.clone(),
@@ -168,13 +181,13 @@ pub async fn run_stream(
             }),
     };
     stream_output.print_final_summary =
-        !args.no_test && should_print_final_summary(continuous, interrupted);
+        !args.no_stats && !args.no_test && should_print_final_summary(continuous, interrupted);
     stream_output.show_running_only_summary_note =
         continuous && interrupted && stream_output.print_final_summary;
     for target in &setup.targets {
-        let target_stats = stats.get(&target.label).expect(
+        let Some(target_stats) = stats.get(&target.label) else {
-            "every prepared target must have a stats collector inserted before the run begins",
+            continue;
-        );
+        };
         if multi_target
             && stream_output.print_final_summary
             && stream_output.config.prints_summary()
@@ -361,17 +374,22 @@ fn process_stdin_event<W: Write>(
     terminal_targets: &mut BoundedTargetSet,
 ) -> io::Result<()> {
     match event {
-        ManagedEvent::TargetStateChanged { target, .. } => {
+        ManagedEvent::TargetStateChanged { target, .. } if stream_output.collect_stats => {
             stats
                 .entry(target)
                 .or_insert_with(|| StatsCollector::new(stats_config(true)));
         }
         ManagedEvent::Client { target, event } => {
-            let collector = stats
+            let update = stream_output.collect_stats.then(|| {
-                .entry(target.clone())
+                let collector = stats
-                .or_insert_with(|| StatsCollector::new(stats_config(true)));
+                    .entry(target.clone())
-            let update = collector.process(&event);
+                    .or_insert_with(|| StatsCollector::new(stats_config(true)));
-            stream_output.print_event(&event, Some(target.id.as_str()), &update)?;
+                if matches!(event, ClientEvent::SessionStarted(_)) {
+                    *collector = StatsCollector::new(stats_config(true));
+                }
+                collector.process(&event)
+            });
+            stream_output.print_event(&event, Some(target.id.as_str()), update.as_ref())?;
         }
         ManagedEvent::TargetFinished { outcome }
             if terminal_targets.insert(outcome.target.clone()) =>
@@ -402,7 +420,11 @@ fn reconcile_stdin_stats(
 fn snapshot_stdin_summary_targets(
     stats: &mut BTreeMap<TargetInstance, StatsCollector>,
     status: &ManagedStatus,
+    enabled: bool,
 ) -> BTreeSet<TargetInstance> {
+    if !enabled {
+        return BTreeSet::new();
+    }
     let targets = status
         .targets
         .iter()
@@ -427,11 +449,19 @@ async fn run_stdin_stream(
     setup: ManagedRunSetup,
     output_config: OutputConfig,
     mut shutdown: tokio::sync::watch::Receiver<bool>,
+    args: &ClientArgs,
 ) -> Result<(), Box<dyn std::error::Error>> {
     if *shutdown.borrow() {
         return Ok(());
     }
-    let (task, handle) = ManagedClient::task(setup.managed_config(), setup.managed_targets())?;
+    let (task, handle) = ManagedClient::task(
+        {
+            let mut config = setup.managed_config();
+            config.reconnect_timeout = args.reconnect.then_some(args.reconnect_timeout);
+            config
+        },
+        setup.managed_targets(),
+    )?;
     let mut events = handle.subscribe()?;
     let mut status = handle.subscribe_status();
     let owner = ManagedWorker::start(task, handle.clone())?;
@@ -440,6 +470,8 @@ async fn run_stdin_stream(
 
     let mut stdout = io::LineWriter::new(io::stdout().lock());
     let mut stream_output = StreamOutput {
+        continuous: args.is_continuous(),
+        collect_stats: !args.no_stats,
         config: output_config,
         header_printed: false,
         print_final_summary: false,
@@ -510,7 +542,11 @@ async fn run_stdin_stream(
             }
             if stop.is_some() {
                 preparing = None;
-                summary_targets = Some(snapshot_stdin_summary_targets(&mut stats, &snapshot));
+                summary_targets = Some(snapshot_stdin_summary_targets(
+                    &mut stats,
+                    &snapshot,
+                    !args.no_stats,
+                ));
                 if request_managed_stop_once(&mut stop_requested) {
                     drop(handle.stop());
                 }
@@ -537,7 +573,7 @@ async fn run_stdin_stream(
                         }
                         Err(error) => {
                             stop = Some(StdinStop::Fatal(format!("invalid --targets-stdin target set: {error}")));
-                            summary_targets = Some(snapshot_stdin_summary_targets(&mut stats, &status.borrow()));
+                            summary_targets = Some(snapshot_stdin_summary_targets(&mut stats, &status.borrow(), !args.no_stats));
                             if request_managed_stop_once(&mut stop_requested) { drop(handle.stop()); }
                         }
                     }
@@ -558,7 +594,7 @@ async fn run_stdin_stream(
                     }
                     Err(_) => {
                         stop = Some(StdinStop::Fatal("--targets-stdin update was rejected".to_owned()));
-                        summary_targets = Some(snapshot_stdin_summary_targets(&mut stats, &status.borrow()));
+                        summary_targets = Some(snapshot_stdin_summary_targets(&mut stats, &status.borrow(), !args.no_stats));
                         if request_managed_stop_once(&mut stop_requested) { drop(handle.stop()); }
                     }
                 }
@@ -595,7 +631,8 @@ async fn run_stdin_stream(
         );
     }
 
-    let print_summary = matches!(stop, Some(StdinStop::Interrupted | StdinStop::Eof));
+    let print_summary =
+        !args.no_stats && matches!(stop, Some(StdinStop::Interrupted | StdinStop::Eof));
     stream_output.print_final_summary = print_summary;
     stream_output.show_running_only_summary_note = print_summary;
     if let Some(summary_targets) = summary_targets {
@@ -636,11 +673,17 @@ fn process_event<W: Write>(
 ) -> io::Result<()> {
     match event {
         ManagedEvent::Client { target, event } => {
-            let collector = stats
+            let update = stream_output.collect_stats.then(|| {
-                .entry(target.id.as_str().to_owned())
+                let collector = stats
-                .or_insert_with(|| StatsCollector::new(stats_config(false)));
+                    .get_mut(target.id.as_str())
-            let update = collector.process(&event);
+                    .expect("prepared target has a statistics collector");
-            stream_output.print_event(&event, Some(target.id.as_str()), &update)?;
+                if matches!(event, ClientEvent::SessionStarted(_)) {
+                    let config = stats_config(stream_output.continuous);
+                    *collector = StatsCollector::new(config);
+                }
+                collector.process(&event)
+            });
+            stream_output.print_event(&event, Some(target.id.as_str()), update.as_ref())?;
         }
         ManagedEvent::TargetFinished { outcome } => {
             terminal_targets.insert(outcome.target.clone());
@@ -678,6 +721,8 @@ fn dropped_event_warning(dropped_events: u64) -> Option<String> {
 }
 
 struct StreamOutput<'a, W: Write> {
+    continuous: bool,
+    collect_stats: bool,
     config: OutputConfig,
     header_printed: bool,
     print_final_summary: bool,
@@ -689,10 +734,10 @@ impl<W: Write> StreamOutput<'_, W> {
         &mut self,
         event: &ClientEvent,
         target: Option<&str>,
-        update: &EventStatsUpdate,
+        update: Option<&EventStatsUpdate>,
     ) -> io::Result<()> {
         self.print_header()?;
-        if let Some(line) = self.config.render_event(event, target, Some(update)) {
+        if let Some(line) = self.config.render_event(event, target, update) {
             writeln!(self.out, "{line}")?;
         }
         Ok(())
@@ -734,7 +779,7 @@ fn stats_config(continuous: bool) -> StatsConfig {
     }
 }
 fn finite_stats_memory_warning(args: &ClientArgs, target_count: usize) -> Option<String> {
-    if args.no_test || args.is_continuous() || args.duration.is_zero() {
+    if args.no_stats || args.no_test || args.is_continuous() || args.duration.is_zero() {
         return None;
     }
     let target_count = u64::try_from(target_count).unwrap_or(u64::MAX);
@@ -761,3 +806,8 @@ fn finite_stats_memory_warning(args: &ClientArgs, target_count: usize) -> Option
     };
     Some(format!("irtt-rs: warning: finite exact statistics may retain about {formatted} for this run; {guidance}"))
 }
+
+fn reconnect_stats_warning(args: &ClientArgs) -> Option<&'static str> {
+    (args.reconnect && !args.no_stats).then_some("irtt-rs: warning: statistics spanning multiple sessions may not accurately represent one uninterrupted measurement run; use --no-stats for measurement events only. Session statistics reset on reconnect.")
+}
+

crates/irtt-client/src/managed/model.rs

@@ -101,6 +101,9 @@ pub struct ManagedClientConfig {
     pub outcome_history_limit: usize,
     pub max_live_target_generations: usize,
     pub final_drain: Duration,
+    /// Retry failed/closed sessions and detect silence while sending. Only valid
+    /// for continuous measurement runs. None (the default) disables reconnection.
+    pub reconnect_timeout: Option<Duration>,
 }
 
 impl Default for ManagedClientConfig {
@@ -114,6 +117,7 @@ impl Default for ManagedClientConfig {
             outcome_history_limit: DEFAULT_MANAGED_OUTCOME_HISTORY_LIMIT,
             max_live_target_generations: DEFAULT_MANAGED_MAX_LIVE_TARGET_GENERATIONS,
             final_drain: DEFAULT_MANAGED_FINAL_DRAIN,
+            reconnect_timeout: None,
         }
     }
 }
@@ -364,6 +368,8 @@ pub struct ManagedTargetFailure {
 /// Configuration failure detected before any runtime work begins.
 #[derive(Debug, Error)]
 pub enum ManagedConfigError {
+    #[error("reconnection requires continuous measurement mode and a positive, schedulable silence timeout")]
+    InvalidReconnectPolicy,
     #[error("FinishWhenQuiescent requires at least one initial target")]
     EmptyInitialTargets,
     #[error("duplicate managed target id {id}")]

crates/irtt-client/src/managed/task.rs

@@ -327,7 +327,14 @@ impl ManagedClientTask {
         let mut non_send_deadline = None;
         let mut send_deadline = None;
         for target in &self.targets {
+            non_send_deadline = non_send_deadline
+                .into_iter()
+                .chain(target.silence_deadline(self.config.reconnect_timeout))
+                .min();
             match target.phase() {
+                TargetPhase::Retrying { deadline } => {
+                    non_send_deadline = non_send_deadline.into_iter().chain(Some(*deadline)).min();
+                }
                 TargetPhase::Active {
                     client,
                     schedule,

crates/irtt-client/src/managed/task/control.rs

@@ -367,6 +367,15 @@ fn build_task(
             maximum: tokio::sync::Semaphore::MAX_PERMITS,
         });
     }
+    if let Some(timeout) = config.reconnect_timeout {
+        if timeout.is_zero()
+            || Instant::now().checked_add(timeout).is_none()
+            || config.client.request.duration.is_some()
+            || config.client.request.run_mode != crate::RunMode::Normal
+        {
+            return Err(ManagedConfigError::InvalidReconnectPolicy);
+        }
+    }
     if config.max_live_target_generations == 0 {
         return Err(ManagedConfigError::ZeroLiveGenerationLimit);
     }
@@ -417,6 +426,10 @@ fn build_task(
             server_addr: Arc::from(target.server_addr),
             remote: None,
             counters: TargetCounters::default(),
+            sent_before_session: 0,
+            retry_delay: std::time::Duration::from_millis(250),
+            silence_since: None,
+            last_reply_at: None,
             phase: Some(TargetPhase::Pending { client_config }),
         });
     }

crates/irtt-client/src/managed/task/pacing.rs

@@ -206,6 +206,12 @@ impl ManagedClientTask {
         };
         self.targets[index].sync_packets_sent(&client);
         send_waiting = result.is_pending();
+        if send_waiting {
+            // A blocked sender is not evidence of an unresponsive peer. Arm
+            // a fresh silence period only after transmission resumes.
+            self.targets[index].silence_since = None;
+            self.targets[index].last_reply_at = None;
+        }
         self.targets[index].restore_phase(TargetPhase::Active {
             client,
             schedule,

crates/irtt-client/src/managed/task/target.rs

@@ -5,7 +5,7 @@ use std::{
     pin::Pin,
     sync::Arc,
     task::{Context, Poll},
-    time::Instant,
+    time::{Duration, Instant},
 };
 
 use crate::{
@@ -27,6 +27,9 @@ type ConnectFuture =
     Pin<Box<dyn Future<Output = Result<AsyncClient, ClientError>> + Send + 'static>>;
 
 pub(super) enum TargetPhase {
+    Retrying {
+        deadline: Instant,
+    },
     Pending {
         client_config: ClientConfig,
     },
@@ -65,6 +68,7 @@ pub(super) enum TargetPhase {
 impl TargetPhase {
     pub(super) fn lifecycle(&self) -> ManagedTargetLifecycle {
         match self {
+            Self::Retrying { .. } => ManagedTargetLifecycle::Pending,
             Self::Pending { .. } => ManagedTargetLifecycle::Pending,
             Self::Connecting { .. } => ManagedTargetLifecycle::Connecting,
             Self::Opening { .. } => ManagedTargetLifecycle::Opening,
@@ -182,11 +186,30 @@ pub(super) struct TargetRuntime {
     pub(super) server_addr: Arc<str>,
     pub(super) remote: Option<std::net::SocketAddr>,
     pub(super) counters: TargetCounters,
+    pub(super) sent_before_session: u64,
+    pub(super) retry_delay: Duration,
+    pub(super) silence_since: Option<Instant>,
+    pub(super) last_reply_at: Option<Instant>,
     // None only while a synchronous operation owns the phase locally.
     pub(super) phase: Option<TargetPhase>,
 }
 
 impl TargetRuntime {
+    pub(super) fn silence_deadline(&self, timeout: Option<Duration>) -> Option<Instant> {
+        let TargetPhase::Active {
+            schedule,
+            send_waiting,
+            ..
+        } = self.phase()
+        else {
+            return None;
+        };
+        if *send_waiting || schedule.is_finished() {
+            return None;
+        }
+        self.silence_since?.checked_add(timeout?)
+    }
+
     pub(super) fn phase(&self) -> &TargetPhase {
         self.phase
             .as_ref()
@@ -218,16 +241,18 @@ impl TargetRuntime {
         self.membership.is_desired() && matches!(phase, TargetPhase::Active { .. })
     }
 
-    /// Adopt the underlying client's authoritative sent count.
+    /// Combine the current client's authoritative sent count with retired
-    ///
+    /// sessions, so one target generation keeps durable accounting on retries.
-    /// The client owns `packets_sent`; this runtime only mirrors it so a
-    /// terminal outcome can report it after the client is gone.
     pub(super) fn sync_packets_sent(&mut self, client: &AsyncClient) {
-        self.counters.packets_sent = client.packets_sent();
+        self.counters.packets_sent = self
+            .sent_before_session
+            .saturating_add(client.packets_sent());
     }
 
     pub(super) fn sync_live_packets_sent(&mut self) {
-        self.counters.packets_sent = self.phase().client().packets_sent();
+        self.counters.packets_sent = self
+            .sent_before_session
+            .saturating_add(self.phase().client().packets_sent());
     }
 
     /// Build this target's durable outcome.
@@ -284,6 +309,30 @@ impl ManagedClientTask {
 
     pub(super) fn publish_client_events(&mut self, index: usize, events: Vec<ClientEvent>) {
         for event in events {
+            match &event {
+                ClientEvent::EchoSent { sent_at, .. } => {
+                    let target = &mut self.targets[index];
+                    let anchor = target
+                        .last_reply_at
+                        .filter(|reply| {
+                            self.config.reconnect_timeout.is_some_and(|timeout| {
+                                sent_at.mono.duration_since(*reply) < timeout
+                            })
+                        })
+                        .unwrap_or(sent_at.mono);
+                    target.silence_since.get_or_insert(anchor);
+                }
+                ClientEvent::EchoReply { received_at, .. }
+                | ClientEvent::LateReply { received_at, .. }
+                | ClientEvent::DuplicateReply { received_at, .. } => {
+                    // A reply disarms silence until another probe is committed.
+                    // A healthy sender with a long interval can remain idle.
+                    self.targets[index].silence_since = None;
+                    self.targets[index].last_reply_at = Some(received_at.mono);
+                    self.targets[index].retry_delay = Duration::from_millis(250);
+                }
+                _ => {}
+            }
             self.targets[index].counters.observe(&event);
             self.publish_event(ManagedEvent::Client {
                 target: self.targets[index].instance.clone(),
@@ -298,6 +347,23 @@ impl ManagedClientTask {
         end_reason: ManagedTargetEndReason,
         cleanup_failure: Option<ManagedTargetFailure>,
     ) {
+        if self.can_reconnect(index)
+            && matches!(
+                end_reason,
+                ManagedTargetEndReason::Failed(_)
+                    | ManagedTargetEndReason::PeerClosed
+                    | ManagedTargetEndReason::TestComplete
+            )
+        {
+            let target = &mut self.targets[index];
+            target.sent_before_session = target.counters.packets_sent;
+            target.silence_since = None;
+            target.last_reply_at = None;
+            let deadline = Instant::now() + target.retry_delay;
+            target.retry_delay = (target.retry_delay * 2).min(Duration::from_secs(10));
+            self.install_target_phase(index, TargetPhase::Retrying { deadline });
+            return;
+        }
         let outcome = Arc::new(self.targets[index].outcome(end_reason, cleanup_failure));
         // Finalize history before the phase publisher installs Terminal, updates
         // pacing and publishes one complete status followed by the lifecycle event.
@@ -311,6 +377,12 @@ impl ManagedClientTask {
         self.publish_event(ManagedEvent::TargetFinished { outcome });
     }
 
+    fn can_reconnect(&self, index: usize) -> bool {
+        self.config.reconnect_timeout.is_some()
+            && !self.phase.is_stopping()
+            && self.targets[index].membership.is_desired()
+    }
+
     fn fail_target(&mut self, index: usize, phase: ManagedTargetFailurePhase, error: ClientError) {
         let failure = classify_client_error(phase, &error);
         self.finish_target(index, ManagedTargetEndReason::Failed(failure), None);
@@ -332,6 +404,9 @@ impl ManagedClientTask {
         };
         client.discard_prepared_probe();
         self.targets[index].sync_live_packets_sent();
+        if self.can_reconnect(index) {
+            return self.begin_close(index, primary_end, None, now);
+        }
         match cleanup {
             OpenSessionFailureCleanup::Drain => self.begin_drain(index, primary_end, now),
             OpenSessionFailureCleanup::Close => self.begin_close(index, primary_end, None, now),
@@ -359,6 +434,27 @@ impl ManagedClientTask {
         let retirement = self.effective_retirement(index);
         let stopping = self.phase.is_stopping() || retirement.is_some();
         match self.targets[index].phase_mut() {
+            TargetPhase::Retrying { deadline } => {
+                if stopping {
+                    self.finish_target(
+                        index,
+                        retirement.unwrap_or(ManagedTargetEndReason::Stopped),
+                        None,
+                    );
+                    false
+                } else if now >= *deadline {
+                    let target = &self.targets[index].config;
+                    let mut client_config = self.config.client.clone();
+                    client_config.address_family = target
+                        .address_family
+                        .unwrap_or(client_config.address_family);
+                    client_config.auth = target.auth.resolve(&client_config.auth);
+                    self.install_target_phase(index, TargetPhase::Pending { client_config });
+                    true
+                } else {
+                    false
+                }
+            }
             TargetPhase::Pending { .. } => {
                 if stopping {
                     self.finish_target(
@@ -465,6 +561,16 @@ impl ManagedClientTask {
                 if stopping {
                     client.discard_prepared_probe();
                     self.targets[index].sync_live_packets_sent();
+                    if self.config.reconnect_timeout.is_some() {
+                        return self.begin_close(
+                            index,
+                            retirement
+                                .clone()
+                                .unwrap_or(ManagedTargetEndReason::Stopped),
+                            None,
+                            now,
+                        );
+                    }
                     return self.begin_drain(
                         index,
                         retirement
@@ -500,6 +606,22 @@ impl ManagedClientTask {
                     self.finish_target(index, ManagedTargetEndReason::PeerClosed, None);
                     return false;
                 }
+                if self.targets[index]
+                    .silence_deadline(self.config.reconnect_timeout)
+                    .is_some_and(|deadline| deadline <= now)
+                {
+                    self.targets[index].sync_live_packets_sent();
+                    return self.begin_close(
+                        index,
+                        ManagedTargetEndReason::Failed(ManagedTargetFailure {
+                            phase: ManagedTargetFailurePhase::Receiving,
+                            kind: ManagedTargetFailureKind::Timeout,
+                            message: Arc::from("session silent while sending probes"),
+                        }),
+                        None,
+                        now,
+                    );
+                }
                 let TargetPhase::Active {
                     client, schedule, ..
                 } = self.targets[index].phase_mut()

crates/irtt-client/src/managed/task/update.rs

@@ -45,7 +45,8 @@ struct UpdatePlan {
 
 fn synchronously_retireable(phase: &TargetPhase) -> bool {
     match phase {
-        TargetPhase::Pending { .. }
+        TargetPhase::Retrying { .. }
+        | TargetPhase::Pending { .. }
         | TargetPhase::Connecting { .. }
         | TargetPhase::Terminal { .. } => true,
         TargetPhase::Opening { open, .. } => !open.has_in_flight_work(),
@@ -140,6 +141,10 @@ impl ManagedClientTask {
                 membership: TargetMembership::Desired,
                 remote: None,
                 counters: TargetCounters::default(),
+                sent_before_session: 0,
+                retry_delay: std::time::Duration::from_millis(250),
+                silence_since: None,
+                last_reply_at: None,
                 phase: Some(TargetPhase::Pending {
                     client_config: planned.client_config,
                 }),

docs/irtt-client.md

@@ -220,6 +220,25 @@ See [MEASUREMENTS AND MEMORY](#measurements-and-memory) below for why this
 distinction also determines what statistics can be computed and how much
 memory a long run retains.
 
+### Optional recovery and event-only runs
+
+`--no-stats` disables statistics collection and final summaries while preserving individual measurement output. Statistics-dependent columns (such as IPDV) are unavailable. This does not remove the `irtt-stats` dependency.
+
+`--reconnect` enables automatic recovery for individual targets in continuous mode (`--continuous` or `--duration 0`). Failed connections, closed sessions, and unresponsive servers are retried without interrupting other targets. Retries use exponential backoff from 250 ms up to 10 seconds.
+
+`--reconnect-timeout DURATION` controls how long a session can go without valid replies before reconnection is attempted (default: `10s`). The timeout applies only while actively sending probes and should account for the probe interval and expected packet loss.
+
+**Statistics reset on reconnection**, so final summaries only reflect the latest session per target. A warning is emitted when reconnection is enabled with statistics collection. Use `--no-stats` to disable collection and suppress the warning.
+
+Example for continuous, multi-target CSV output without statistics:
+
+```sh
+irtt-client --continuous --reconnect --no-stats \
+  --format csv --header never \
+  --columns client_receive_wall_ns,target,seq,rd_us,sd_us \
+  reflector1 reflector2
+```
+
 ## OUTPUT
 
 ### `--format FORMAT`
@@ -355,7 +374,7 @@ approaches a fixed multiple of `targets` regardless of how long it runs.
 ## PEER CLOSE AND EXIT BEHAVIOR
 
 A peer closing a finite run is accepted as a terminal, successful outcome.
-In continuous mode, an uninterrupted peer closure is treated as an error —
+Without `--reconnect`, in continuous mode, an uninterrupted peer closure is treated as an error —
 the client exits nonzero — so a process supervisor can restart it; an
 interrupted run (`Ctrl-C`) never reports peer closure as an error, whichever
 mode it was in.