feat(worker): Handler-Session nach Usage-Limit-Reset automatisch neu starten

Eine List-Handler-ConPTY-Session, die mitten in einer Phase stirbt, war bisher
ein stilles totes Tile: InteractiveTerminalViewModel.OnSessionProcessExited
setzte nur bei einem Tod innerhalb des 5s-Grace-Fensters ein sichtbares Banner.

Neu:
- HandlerRoundState (JSON unter ~/.claudeDo/handler-rounds.json, keine
  EF-Migration): Phase + Task-Set der laufenden Runde, geschrieben von
  BuildForMergeHelperAsync ("triage") und BuildForMergeHelperHandoffAsync.
  Ohne das existieren beide nur im transienten HandoffRequested-Broadcast.
- HandlerRestartCoordinator: entscheidet den Neustart, fail-closed. Beleg fuer
  "war das Usage-Limit" ist FiveHour.Utilization >= 95 zum Todeszeitpunkt --
  eine ConPTY-Session hat kein stdout, TaskRunner.IsUsageLimitMessage hat hier
  nichts zu lesen. Nur bei PowerUse (beide Keys), max ein Neustart pro Runde,
  Warten bis FiveHour.ResetsAt, dann HandoffRequested erneut broadcasten.
  Verbraucht kein Handoff-Budget (HandoffRoundTracker.Register wird nicht
  angefasst). WorkerLog-Zeilen beim Tod und beim Neustart.
- Ui meldet den Exit-Code (nur dort bekannt) via
  WorkerHub.ReportInteractiveSessionExited, ohne selbst zu filtern.

Sibling zu UsageLimitAutoContinueCoordinator, kein Umbau davon: der sieht nur
headless Runs.

src/ClaudeDo.Ui/Services/Interfaces/IWorkerClient.cs

@@ -146,6 +146,10 @@ public interface IWorkerClient : INotifyPropertyChanged
     /// session's role: "wait" | "merge" | "wait_final" | "merge_final".</summary>
     Task<LaunchSpec> GetMergeHelperHandoffLaunchSpecAsync(
         string taskId, IReadOnlyList<string> survivingTaskIds, string nextPhase, CancellationToken ct = default);
+    /// <summary>Reports that a task-based ConPTY session's process exited non-zero after it had been
+    /// running. The exit code only exists in this process; the worker decides whether it was a
+    /// list-handler session killed by the usage limit and reopens the same phase after the 5h reset.</summary>
+    Task ReportInteractiveSessionExitedAsync(string taskId, int exitCode, CancellationToken ct = default);
     /// <summary>Starts a planning session and returns the launch spec for an embedded ConPTY
     /// planning terminal (replaces StartPlanningSessionAsync's external wt window).</summary>
     Task<LaunchSpec> GetPlanningStartLaunchSpecAsync(string taskId, CancellationToken ct = default);

src/ClaudeDo.Ui/Services/WorkerClient.cs

@@ -626,6 +626,9 @@ public partial class WorkerClient : ObservableObject, IAsyncDisposable, IWorkerC
         string taskId, IReadOnlyList<string> survivingTaskIds, string nextPhase, CancellationToken ct = default)
         => await InvokeTimedAsync<LaunchSpec>("GetMergeHelperHandoffLaunchSpec", () => _hub.InvokeAsync<LaunchSpec>("GetMergeHelperHandoffLaunchSpec", taskId, survivingTaskIds, nextPhase, ct));
 
+    public Task ReportInteractiveSessionExitedAsync(string taskId, int exitCode, CancellationToken ct = default)
+        => InvokeTimedAsync("ReportInteractiveSessionExited", () => _hub.InvokeAsync("ReportInteractiveSessionExited", taskId, exitCode, ct));
+
     public async Task<LaunchSpec> GetPlanningStartLaunchSpecAsync(string taskId, CancellationToken ct = default)
         => await InvokeTimedAsync<LaunchSpec>("GetPlanningStartLaunchSpec", () => _hub.InvokeAsync<LaunchSpec>("GetPlanningStartLaunchSpec", taskId, ct));
 

src/ClaudeDo.Ui/ViewModels/InteractiveTerminalViewModel.cs

@@ -96,6 +96,13 @@ public sealed partial class InteractiveTerminalViewModel : ViewModelBase, IDispo
         }
     }
 
+    /// <summary>Raised when the process died non-zero *after* clearing <see cref="StartupGraceWindow"/> —
+    /// i.e. a session that ran, then broke. That used to be a silent dead tile with no signal at all;
+    /// the worker turns it into a usage-limit auto-restart for list-handler sessions
+    /// (<c>HandlerRestartCoordinator</c>). A startup death keeps using <see cref="StartError"/> instead:
+    /// it never reached a state worth resuming.</summary>
+    public event Action<int>? ExitedAfterStartup;
+
     /// <summary>Internal (not private) so tests can drive it directly without a real ConPTY spawn.</summary>
     internal void OnSessionProcessExited(object? sender, int exitCode)
     {
@@ -103,11 +110,16 @@ public sealed partial class InteractiveTerminalViewModel : ViewModelBase, IDispo
         HasExited = true;
         ExitCode = exitCode;
 
+        if (exitCode == 0) return;
+        if (_startedAtUtc is not { } startedAt) return; // never reached "running" — nothing ran to resume
+
         // Died at startup: route through the same banner as a launch-time failure instead of
         // leaving a dead terminal with no Retry affordance (ConPtyPaneViewModel.CanRetry
         // requires StartError).
-        if (exitCode != 0 && _startedAtUtc is { } startedAt && _utcNow() - startedAt < StartupGraceWindow)
+        if (_utcNow() - startedAt < StartupGraceWindow)
             StartError = $"Session exited immediately (code {exitCode}).";
+        else
+            ExitedAfterStartup?.Invoke(exitCode);
     }
 
     /// <summary>Reports a failure that happened before <see cref="Start"/> could be called (e.g. the

src/ClaudeDo.Ui/ViewModels/MissionControl/ConPtyPaneViewModel.cs

@@ -36,6 +36,12 @@ public sealed partial class ConPtyPaneViewModel : ViewModelBase, IDisposable
     /// commits the worktree and moves the task to WaitingForReview.</summary>
     public event Action<string>? SubmitForReviewRequested;
 
+    /// <summary>Raised with <c>(taskId, exitCode)</c> when a task-based session that had been running
+    /// dies non-zero. The host reports it to the worker, which decides whether this was a list
+    /// handler killed by the usage limit and worth reopening. Ad-hoc panes never raise it — there is
+    /// no task to reopen.</summary>
+    public event Action<string, int>? SessionDiedAfterStartup;
+
     /// <summary>Task-based pane — dedup'd by <see cref="TaskId"/>. Pass null for an ad-hoc pane
     /// (no task, never deduped); prefer <see cref="CreateAdHoc"/> at ad-hoc call sites.
     /// <paramref name="descriptorFactory"/> is resolved by the pane itself (in <see cref="Start"/>)
@@ -49,7 +55,34 @@ public sealed partial class ConPtyPaneViewModel : ViewModelBase, IDisposable
         TaskId = taskId;
         _displayTitle = displayTitle;
         _descriptorFactory = descriptorFactory;
-        Terminal.PropertyChanged += OnTerminalPropertyChanged;
+        Subscribe(Terminal);
+    }
+
+    // Terminal is swapped wholesale on Retry, so the handlers follow the property rather than being
+    // wired once at construction. (The initial value comes from the field initializer, which never
+    // runs through the setter -- hence the explicit Subscribe in the constructor.)
+    partial void OnTerminalChanging(InteractiveTerminalViewModel? oldValue, InteractiveTerminalViewModel newValue)
+    {
+        if (oldValue is not null) Unsubscribe(oldValue);
+    }
+
+    partial void OnTerminalChanged(InteractiveTerminalViewModel value) => Subscribe(value);
+
+    private void Subscribe(InteractiveTerminalViewModel terminal)
+    {
+        terminal.PropertyChanged += OnTerminalPropertyChanged;
+        terminal.ExitedAfterStartup += OnTerminalExitedAfterStartup;
+    }
+
+    private void Unsubscribe(InteractiveTerminalViewModel terminal)
+    {
+        terminal.PropertyChanged -= OnTerminalPropertyChanged;
+        terminal.ExitedAfterStartup -= OnTerminalExitedAfterStartup;
+    }
+
+    private void OnTerminalExitedAfterStartup(int exitCode)
+    {
+        if (TaskId is { } id) SessionDiedAfterStartup?.Invoke(id, exitCode);
     }
 
     private readonly Func<System.Threading.Tasks.Task<TerminalLaunchDescriptor>> _descriptorFactory;
@@ -111,16 +144,15 @@ public sealed partial class ConPtyPaneViewModel : ViewModelBase, IDisposable
     [RelayCommand(CanExecute = nameof(CanRetry))]
     private void Retry()
     {
-        Terminal.PropertyChanged -= OnTerminalPropertyChanged;
-        Terminal.Dispose();
-        Terminal = new InteractiveTerminalViewModel();
-        Terminal.PropertyChanged += OnTerminalPropertyChanged;
+        var dead = Terminal;
+        Terminal = new InteractiveTerminalViewModel(); // OnTerminalChanging/Changed move the handlers
+        dead.Dispose();
         Start();
     }
 
     public void Dispose()
     {
-        Terminal.PropertyChanged -= OnTerminalPropertyChanged;
+        Unsubscribe(Terminal);
         Terminal.Kill();
         Terminal.Dispose();
     }

src/ClaudeDo.Ui/ViewModels/MissionControlViewModel.cs

@@ -413,10 +413,28 @@ public sealed partial class MissionControlViewModel : ViewModelBase, IDisposable
         pane.ErrorReported += OnConPtyPaneError;
         pane.CloseRequested += CloseConPtySession;
         pane.SubmitForReviewRequested += OnPaneSubmitForReview;
+        pane.SessionDiedAfterStartup += OnPaneSessionDied;
         ConPtySessions.Add(pane);
         pane.Start();
     }
 
+    // A ConPTY session that had been running died. The UI process owns the terminal, so this is the
+    // only place the exit code exists -- report it and let the worker decide whether it was a list
+    // handler killed by the usage limit and worth reopening after the reset
+    // (HandlerRestartCoordinator). Every other pane is just a no-op there, so no filtering here.
+    private async void OnPaneSessionDied(string taskId, int exitCode)
+    {
+        try
+        {
+            await _worker.ReportInteractiveSessionExitedAsync(taskId, exitCode);
+        }
+        catch (Exception ex)
+        {
+            // A worker that is down cannot restart anything anyway; the dead tile is already visible.
+            ErrorReported?.Invoke(Loc.T("missionControl.conptyLaunchFailed", ex.Message));
+        }
+    }
+
     private static async System.Threading.Tasks.Task<TerminalLaunchDescriptor> DescribeAsync(
         Func<System.Threading.Tasks.Task<LaunchSpec>> fetch)
     {
@@ -457,6 +475,7 @@ public sealed partial class MissionControlViewModel : ViewModelBase, IDisposable
         pane.ErrorReported -= OnConPtyPaneError;
         pane.CloseRequested -= CloseConPtySession;
         pane.SubmitForReviewRequested -= OnPaneSubmitForReview;
+        pane.SessionDiedAfterStartup -= OnPaneSessionDied;
         ConPtySessions.Remove(pane);
         pane.Dispose();
     }
@@ -507,6 +526,7 @@ public sealed partial class MissionControlViewModel : ViewModelBase, IDisposable
             c.ErrorReported -= OnConPtyPaneError;
             c.CloseRequested -= CloseConPtySession;
             c.SubmitForReviewRequested -= OnPaneSubmitForReview;
+            c.SessionDiedAfterStartup -= OnPaneSessionDied;
             c.Dispose();
         }
         ConPtySessions.Clear();

src/ClaudeDo.Worker/Hub/WorkerHub.cs

@@ -72,6 +72,7 @@ public sealed class WorkerHub : Microsoft.AspNetCore.SignalR.Hub
     private readonly TicketPatStore _ticketPat;
     private readonly TicketClientFactory _ticketClients;
     private readonly TicketImportService _ticketImport;
+    private readonly HandlerRestartCoordinator? _handlerRestart;
 
     public WorkerHub(
         QueueService queue,
@@ -110,7 +111,8 @@ public sealed class WorkerHub : Microsoft.AspNetCore.SignalR.Hub
         TicketSystemConfig ticketConfig = null!,
         TicketPatStore ticketPat = null!,
         TicketClientFactory ticketClients = null!,
-        TicketImportService ticketImport = null!)
+        TicketImportService ticketImport = null!,
+        HandlerRestartCoordinator? handlerRestart = null)
     {
         _queue = queue;
         _waker = waker;
@@ -149,6 +151,7 @@ public sealed class WorkerHub : Microsoft.AspNetCore.SignalR.Hub
         _ticketPat = ticketPat;
         _ticketClients = ticketClients;
         _ticketImport = ticketImport;
+        _handlerRestart = handlerRestart;
     }
 
     // Persistence boundary for the session_skills JSON-array columns (task/list/global).
@@ -786,6 +789,18 @@ public sealed class WorkerHub : Microsoft.AspNetCore.SignalR.Hub
         return _interactiveLaunchSpec.BuildForMergeHelperHandoffAsync(taskId, survivingTaskIds, nextPhase, Context.ConnectionAborted);
     });
 
+    // Reports that an embedded ConPTY session's process exited. Only the UI process owns the
+    // terminal, so the exit code exists nowhere else; every decision about it (was this a list
+    // handler, did the usage limit kill it, is PowerUse on, when does the window reset) belongs to
+    // HandlerRestartCoordinator. Fire-and-forget on purpose: the coordinator's task stays parked
+    // until the 5h reset, which is far longer than this connection or this hub instance lives, so
+    // it must not run under Context.ConnectionAborted either.
+    public void ReportInteractiveSessionExited(string taskId, int exitCode)
+    {
+        if (_handlerRestart is null) return;
+        _ = _handlerRestart.OnSessionExitedAsync(taskId, exitCode, CancellationToken.None);
+    }
+
     // Starts a planning session (worktree + prompt files + token, task -> Planning) and returns
     // the launch spec for an embedded ConPTY planning terminal -- the ConPTY replacement for
     // StartPlanningSessionAsync's external wt window. On any spec-build failure the just-started

src/ClaudeDo.Worker/Program.cs

@@ -95,6 +95,8 @@ builder.Services.AddSingleton<IRepoCloner, GitRepoCloner>();
 builder.Services.AddSingleton<ISessionSkillRegistry, SessionSkillRegistry>();
 builder.Services.AddSingleton<ISessionSkillSeeder, SessionSkillSeeder>();
 builder.Services.AddSingleton<HandoffRoundTracker>();
+builder.Services.AddSingleton<HandlerRoundState>();
+builder.Services.AddSingleton<HandlerRestartCoordinator>();
 builder.Services.AddSingleton<InteractiveLaunchSpecService>();
 builder.Services.AddSingleton<TaskRunner>();
 builder.Services.AddSingleton<WorktreeMaintenanceService>();

src/ClaudeDo.Worker/Runner/HandlerRestartCoordinator.cs

@@ -0,0 +1,129 @@
+using ClaudeDo.Data;
+using ClaudeDo.Data.Models;
+using ClaudeDo.Data.Repositories;
+using ClaudeDo.Worker.Hub;
+using ClaudeDo.Worker.Usage;
+using Microsoft.EntityFrameworkCore;
+
+namespace ClaudeDo.Worker.Runner;
+
+/// <summary>
+/// Reopens a list-handler ConPTY session that died on the Anthropic usage limit, once the 5h window
+/// resets — the interactive counterpart to <c>UsageLimitAutoContinueCoordinator</c>, which only ever
+/// sees headless task runs. Deliberately a sibling and not an extension of it: a ConPTY session has
+/// no <c>task_runs</c> row, no <c>FailureReason</c>, and is resumed by relaunching a phase rather
+/// than by <c>ContinueTask</c> through the override slot.
+///
+/// Driven by the UI (<c>WorkerHub.ReportInteractiveSessionExited</c>) because the exit code only
+/// exists in the process that owns the terminal. Everything else — was it the usage limit, is
+/// PowerUse on, which phase and task set, when does the window reset — is decided here.
+///
+/// Fails CLOSED, unlike the usage gate: every missing signal means "don't relaunch". An interactive
+/// session that pops back up at 3am uninvited is worse than one that stays dead.
+/// </summary>
+public sealed class HandlerRestartCoordinator
+{
+    // The corroborating signal that this exit really was the usage limit. A ConPTY session gives us
+    // nothing but an exit code -- there is no stdout to match against, so TaskRunner's
+    // IsUsageLimitMessage regex (which reads the CLI's own "session limit reached" text) has nothing
+    // to read here. The 5h utilization at the moment of death is the one independent witness we do
+    // have, and it is pinned at 100% when the limit is what stopped the session. A crash, a manual
+    // Ctrl-C, or a bad launch at 40% utilization is therefore NOT restarted.
+    internal const int UsageLimitProofPct = 95;
+
+    private readonly IDbContextFactory<ClaudeDoDbContext> _dbFactory;
+    private readonly UsageState _usageState;
+    private readonly HandlerRoundState _rounds;
+    private readonly HubBroadcaster _broadcaster;
+    private readonly ILogger<HandlerRestartCoordinator> _logger;
+
+    public HandlerRestartCoordinator(
+        IDbContextFactory<ClaudeDoDbContext> dbFactory,
+        UsageState usageState,
+        HandlerRoundState rounds,
+        HubBroadcaster broadcaster,
+        ILogger<HandlerRestartCoordinator> logger)
+    {
+        _dbFactory = dbFactory;
+        _usageState = usageState;
+        _rounds = rounds;
+        _broadcaster = broadcaster;
+        _logger = logger;
+    }
+
+    /// <summary>Test seams — a real 5h wait is not a test.</summary>
+    internal Func<DateTimeOffset> NowUtc { get; set; } = () => DateTimeOffset.UtcNow;
+    internal Func<TimeSpan, CancellationToken, Task> Delay { get; set; } = (d, ct) => Task.Delay(d, ct);
+
+    /// <summary>Handles a ConPTY session's process exit. Returns true when a restart was claimed and
+    /// carried out (the returned task completes only after the wait, so a caller that wants
+    /// fire-and-forget must not await it). Any non-handler task, clean exit, non-usage-limit death,
+    /// PowerUse-off list, or already-used round is a silent no-op.</summary>
+    public async Task<bool> OnSessionExitedAsync(string taskId, int exitCode, CancellationToken ct = default)
+    {
+        if (exitCode == 0 || string.IsNullOrEmpty(taskId)) return false;
+
+        var round = _rounds.Get(taskId);
+        // Not a list-handler session at all (a plain interactive task pane), or the initial triage
+        // session, which has no handoff launch spec to reopen -- BuildForMergeHelperAsync needs the
+        // list id and the original selection, and its dedupe/enhance work is not idempotent.
+        if (round is null || !MergeHelperPhase.All.Contains(round.Phase, StringComparer.Ordinal)) return false;
+
+        var snapshot = _usageState.Snapshot;
+        if (snapshot?.FiveHour is not { } fiveHour) return false;
+        if (fiveHour.Utilization < UsageLimitProofPct) return false;
+        if (fiveHour.ResetsAt is not { } resetsAt) return false;
+
+        int number;
+        try
+        {
+            await using var ctx = await _dbFactory.CreateDbContextAsync(ct);
+            var task = await new TaskRepository(ctx).GetByIdAsync(taskId, ct);
+            if (task is null) return false;
+            number = task.Number;
+
+            // Both PowerUse keys, exactly as ComposeHandlerPromptAsync gates the PowerUse prompt
+            // addendum: the global switch alone must not make an interactive run on some unrelated
+            // list restart itself overnight.
+            if (!(await new AppSettingsRepository(ctx).GetAsync(ct)).PowerUseEnabled) return false;
+            if ((await new ListRepository(ctx).GetConfigAsync(task.ListId, ct))?.PowerUseAllowed != true) return false;
+        }
+        catch (Exception ex)
+        {
+            _logger.LogWarning(ex, "HandlerRestartCoordinator: could not evaluate handler task {TaskId}; not restarting", taskId);
+            return false;
+        }
+
+        // One restart per round, claimed before the wait -- a restart that dies again lands back
+        // here and must stop, not relaunch.
+        if (!_rounds.TryClaimAutoRestart(taskId)) return false;
+
+        try
+        {
+            await _broadcaster.WorkerLog(
+                $"List handler #{number} died at the usage limit (5h {fiveHour.Utilization:0}%) — " +
+                $"restarting phase {round.Phase} after the reset at {resetsAt.ToLocalTime():HH:mm}",
+                WorkerLogLevel.Warn, DateTime.UtcNow);
+
+            var wait = resetsAt - NowUtc();
+            if (wait > TimeSpan.Zero) await Delay(wait, ct);
+
+            await _broadcaster.HandoffRequested(taskId, round.TaskIds, round.Phase);
+            await _broadcaster.WorkerLog(
+                $"Auto-restarting list handler #{number} at phase {round.Phase} after usage-limit reset",
+                WorkerLogLevel.Info, DateTime.UtcNow);
+            return true;
+        }
+        catch (OperationCanceledException)
+        {
+            return false;
+        }
+        catch (Exception ex)
+        {
+            // Callers fire-and-forget this (WorkerHub.ReportInteractiveSessionExited), so an
+            // unobserved throw here would be invisible.
+            _logger.LogError(ex, "HandlerRestartCoordinator: auto-restart of handler task {TaskId} failed", taskId);
+            return false;
+        }
+    }
+}

src/ClaudeDo.Worker/Runner/HandlerRoundState.cs

@@ -0,0 +1,138 @@
+using System.Text.Json;
+using ClaudeDo.Data;
+
+namespace ClaudeDo.Worker.Runner;
+
+/// <summary>One list-handler round: the phase a session is running and the exact task set it was
+/// handed. Written every time a handler session is built, so an outside observer (a crashed tile,
+/// a deferred handoff) can reopen the SAME phase with the SAME set without the in-flight
+/// <c>HandoffRequested</c> broadcast that is the only other place these ids exist.</summary>
+/// <param name="Phase">A <see cref="MergeHelperPhase"/> value, or "triage" for the initial session
+/// (which has no handoff spec and therefore cannot be reopened — see HandlerRestartCoordinator).</param>
+/// <param name="AutoRestartUsed">Claimed once per round by HandlerRestartCoordinator, so a handler
+/// that dies again after its auto-restart stays dead instead of relaunching all night.</param>
+/// <param name="DeferredHandoffAt">Reserved for the "hand off cleanly before the usage limit"
+/// follow-up, which parks a handoff until the 5h reset and needs to survive a worker restart.</param>
+public sealed record HandlerRound(
+    string HandlerTaskId,
+    string Phase,
+    IReadOnlyList<string> TaskIds,
+    DateTime StartedAtUtc,
+    bool AutoRestartUsed = false,
+    DateTimeOffset? DeferredHandoffAt = null);
+
+/// <summary>
+/// File-backed store for <see cref="HandlerRound"/>, one entry per handler task. Deliberately a
+/// JSON file under <c>~/.claudeDo/</c> (same Load/Save shape as <c>ClaudeDo.Ui.AppSettings</c>)
+/// rather than a DB column: rounds are process-lifetime scratch state that no query joins against,
+/// and an EF migration here would collide with the one the settings work owns this round.
+/// Every read/write goes through one lock and re-serializes the whole file — there are at most a
+/// handful of concurrent handler runs, so nothing here needs to be clever.
+/// </summary>
+public sealed class HandlerRoundState
+{
+    // Entries older than this are dropped on the next save. A round that has been open for a week
+    // is not coming back, and without a sweep the file only ever grows.
+    private static readonly TimeSpan Retention = TimeSpan.FromDays(7);
+
+    private readonly string _path;
+    private readonly object _lock = new();
+    private Dictionary<string, HandlerRound>? _rounds;
+
+    public HandlerRoundState() : this(Path.Combine(Paths.AppDataRoot(), "handler-rounds.json"))
+    {
+    }
+
+    /// <summary>Test seam — point the store at a temp file instead of the live app data root.</summary>
+    internal HandlerRoundState(string path) => _path = path;
+
+    public HandlerRound? Get(string handlerTaskId)
+    {
+        lock (_lock)
+            return Rounds().GetValueOrDefault(handlerTaskId);
+    }
+
+    /// <summary>Overwrites this handler task's round. <see cref="HandlerRound.AutoRestartUsed"/> is
+    /// carried over when the phase is unchanged: the auto-restart itself rebuilds the same phase's
+    /// launch spec, which lands right back here, and a naive overwrite would hand the round a fresh
+    /// restart budget every time — exactly the all-night relaunch loop the claim exists to prevent.
+    /// A genuinely new phase is a new round and starts with a fresh budget.</summary>
+    public void Save(HandlerRound round)
+    {
+        lock (_lock)
+        {
+            var rounds = Rounds();
+            if (rounds.TryGetValue(round.HandlerTaskId, out var previous)
+                && string.Equals(previous.Phase, round.Phase, StringComparison.Ordinal)
+                && previous.AutoRestartUsed)
+            {
+                round = round with { AutoRestartUsed = true };
+            }
+
+            rounds[round.HandlerTaskId] = round;
+            Persist(rounds);
+        }
+    }
+
+    /// <summary>Takes this round's single auto-restart budget. False when there is no round for the
+    /// task or its budget is already spent — the same "never twice" discipline
+    /// <c>TaskRepository.TryClaimUsageLimitAutoContinueAsync</c> enforces for headless runs.</summary>
+    public bool TryClaimAutoRestart(string handlerTaskId)
+    {
+        lock (_lock)
+        {
+            var rounds = Rounds();
+            if (!rounds.TryGetValue(handlerTaskId, out var round) || round.AutoRestartUsed) return false;
+
+            rounds[handlerTaskId] = round with { AutoRestartUsed = true };
+            Persist(rounds);
+            return true;
+        }
+    }
+
+    public void Remove(string handlerTaskId)
+    {
+        lock (_lock)
+        {
+            var rounds = Rounds();
+            if (rounds.Remove(handlerTaskId)) Persist(rounds);
+        }
+    }
+
+    // Callers hold _lock.
+    private Dictionary<string, HandlerRound> Rounds()
+    {
+        if (_rounds is not null) return _rounds;
+        try
+        {
+            if (File.Exists(_path))
+            {
+                _rounds = JsonSerializer.Deserialize<Dictionary<string, HandlerRound>>(
+                    File.ReadAllText(_path), new JsonSerializerOptions { PropertyNameCaseInsensitive = true });
+            }
+        }
+        catch
+        {
+            // A corrupt/unreadable file must never block a handler launch — start empty.
+        }
+        return _rounds ??= new Dictionary<string, HandlerRound>(StringComparer.Ordinal);
+    }
+
+    private void Persist(Dictionary<string, HandlerRound> rounds)
+    {
+        var cutoff = DateTime.UtcNow - Retention;
+        foreach (var stale in rounds.Where(kv => kv.Value.StartedAtUtc < cutoff).Select(kv => kv.Key).ToList())
+            rounds.Remove(stale);
+
+        try
+        {
+            var dir = Path.GetDirectoryName(_path);
+            if (!string.IsNullOrEmpty(dir)) Directory.CreateDirectory(dir);
+            File.WriteAllText(_path, JsonSerializer.Serialize(rounds, new JsonSerializerOptions { WriteIndented = true }));
+        }
+        catch
+        {
+            // Losing persistence costs an auto-restart across a worker restart, never a launch.
+        }
+    }
+}

src/ClaudeDo.Worker/Runner/InteractiveLaunchSpecService.cs

@@ -44,6 +44,7 @@ public sealed class InteractiveLaunchSpecService
     private readonly GitService _git;
     private readonly string _claudePath;
     private readonly HandoffRoundTracker _handoffRounds;
+    private readonly HandlerRoundState _roundState;
 
     public InteractiveLaunchSpecService(
         IDbContextFactory<ClaudeDoDbContext> dbFactory,
@@ -52,7 +53,8 @@ public sealed class InteractiveLaunchSpecService
         WorktreeManager wtManager,
         GitService git,
         WorkerConfig cfg,
-        HandoffRoundTracker handoffRounds)
+        HandoffRoundTracker handoffRounds,
+        HandlerRoundState roundState)
     {
         _dbFactory = dbFactory;
         _skillSeeder = skillSeeder;
@@ -61,6 +63,7 @@ public sealed class InteractiveLaunchSpecService
         _git = git;
         _claudePath = cfg.ClaudeBin;
         _handoffRounds = handoffRounds;
+        _roundState = roundState;
     }
 
     /// <summary>Builds a LaunchSpec for opening an interactive session in a task's worktree.
@@ -342,9 +345,20 @@ public sealed class InteractiveLaunchSpecService
             ["MCP_TOOL_TIMEOUT"] = McpToolTimeoutMs,
         };
 
+        // The opening round. Recorded so anything watching a handler run knows which round is live
+        // (the deferred-handoff follow-up needs that), but marked "triage" rather than a
+        // MergeHelperPhase value on purpose: there is no handoff spec for this phase, so
+        // HandlerRestartCoordinator will not reopen it.
+        _roundState.Save(new HandlerRound(handlerTaskId, TriagePhase, taskIds, DateTime.UtcNow));
+
         return new LaunchSpec(repoDir, resolvedClaude, args, env);
     }
 
+    /// <summary>The <see cref="HandlerRound.Phase"/> marker for the initial (dedupe/enhance) handler
+    /// session. Deliberately not a <see cref="MergeHelperPhase"/> value — that set is exactly the
+    /// phases <see cref="BuildForMergeHelperHandoffAsync"/> can rebuild.</summary>
+    public const string TriagePhase = "triage";
+
     // Builds the LaunchSpec for the fresh session a merge-helper run hands off to for the given
     // phase -- SAME handler task id as the run that called handoff_list_handler, so
     // HandlerBaseCommit/HandlerHeadCommit and the review range stay untouched; this never creates
@@ -449,6 +463,11 @@ public sealed class InteractiveLaunchSpecService
             ["MCP_TOOL_TIMEOUT"] = McpToolTimeoutMs,
         };
 
+        // Persist the round that is about to start. Phase + task set only exist in the transient
+        // HandoffRequested broadcast otherwise, so without this a dead tile cannot be reopened at
+        // the same phase with the same set (see HandlerRestartCoordinator).
+        _roundState.Save(new HandlerRound(taskId, nextPhase, survivingTaskIds, DateTime.UtcNow));
+
         return new LaunchSpec(repoDir, resolvedClaude, args, env);
     }