chore(claude-do): merge fix(worker): Stuck-Running-Fenster in Continue- und Stop-Pfa

ClaudeDo-Task: 6db1a9e2-aa1d-4911-a763-76b136f97946

src/ClaudeDo.Worker/Queue/QueueService.cs

@@ -348,9 +348,32 @@ public sealed class QueueService : BackgroundService
         }
         catch (OperationCanceledException)
         {
-            // Cancellation is driven by the cancel path, which already wrote the terminal status.
+            // Most cancellation origins (TaskStateService.CancelAsync, the runner's own
-            // Marking the task Failed here would be a regression (it would stomp Cancelled).
+            // OCE handling) already write a terminal status before the CTS is actually
+            // cancelled, so by the time we get here the task has already left Running.
+            // A plain Stop (QueueService.CancelTask cancels the slot's CTS directly, with no
+            // TaskStateService write) landing during the pre-dispatch reads above is the one
+            // gap nothing else covers — the picker's claim already committed status='running'
+            // and nothing since has moved it off that. Only close that specific gap: write
+            // Cancelled when (and only when) the task is still Running, so an origin that
+            // already wrote its own terminal status is never stomped.
             _logger.LogInformation("Slot runner cancelled for task {TaskId}", taskId);
+            try
+            {
+                TaskStatus current;
+                using (var context = _dbFactory.CreateDbContext())
+                    current = await context.Tasks.AsNoTracking()
+                        .Where(t => t.Id == taskId)
+                        .Select(t => t.Status)
+                        .FirstAsync(CancellationToken.None);
+
+                if (current == TaskStatus.Running)
+                    await _state.CancelAsync(taskId, DateTime.UtcNow, CancellationToken.None);
+            }
+            catch (Exception cancelEx)
+            {
+                _logger.LogError(cancelEx, "Could not finalize cancelled task {TaskId} after slot cancellation", taskId);
+            }
         }
         catch (Exception ex)
         {

src/ClaudeDo.Worker/Runner/TaskRunTokenRegistry.cs

@@ -17,6 +17,11 @@ public sealed class TaskRunTokenRegistry
         return false;
     }
     public void Unregister(string token) => _tokenToTaskId.TryRemove(token, out _);
+
+    // Test-only leak check: a run must never leave its token registered after it ends,
+    // no matter which step of setup failed.
+    public bool HasAnyTokenFor(string taskId) => _tokenToTaskId.Values.Contains(taskId);
+
     public static string GenerateToken()
     {
         var bytes = RandomNumberGenerator.GetBytes(32);

src/ClaudeDo.Worker/Runner/TaskRunner.cs

@@ -124,7 +124,7 @@ public sealed class TaskRunner
             var runDir = prep.RunDir!;
 
             var resolvedConfig = await ResolveConfigAsync(task, list, listConfig, null, ct);
-            (mcpToken, mcpConfigPath, resolvedConfig) = await SetupMcpConfigAsync(task, resolvedConfig, ct);
+            (mcpToken, mcpConfigPath, resolvedConfig) = await SetupMcpConfigAsync(task, resolvedConfig, ct, token => mcpToken = token);
 
             await _skillSeeder.SeedAsync(runDir, resolvedConfig.SkillNames, wtCtx is not null, ct);
 
@@ -237,32 +237,35 @@ public sealed class TaskRunner
             runDir = Path.Combine(_cfg.SandboxRoot, taskId);
         }
 
-        var now = DateTime.UtcNow;
+        string? mcpToken = null;
-        // See RunAsync: queue dispatches arrive pre-claimed by the picker.
+        string? mcpConfigPath = null;
-        if (!alreadyClaimed)
-        {
-            var startResult = await _state.StartRunningAsync(taskId, now, ct);
-            if (!startResult.Ok)
-            {
-                _logger.LogWarning("Task {TaskId} skipped: StartRunningAsync rejected ({Reason})", taskId, startResult.Reason);
-                return;
-            }
-        }
-        else
-        {
-            // Queue-claimed dispatches skip StartRunningAsync, so broadcast TaskUpdated here
-            // (see RunAsync for the full rationale).
-            await _broadcaster.TaskUpdated(taskId);
-        }
-        await _broadcaster.TaskStarted(slot, taskId, now);
-
-        await _skillSeeder.SeedAsync(runDir, resolvedConfig.SkillNames, wtCtx is not null, ct);
-
-        var (mcpToken, mcpConfigPath, mcpConfig) = await SetupMcpConfigAsync(task, resolvedConfig, ct);
-        resolvedConfig = mcpConfig;
-
         try
         {
+            var now = DateTime.UtcNow;
+            // See RunAsync: queue dispatches arrive pre-claimed by the picker. Claim, skill
+            // seeding, and MCP setup all live inside this try (mirroring RunAsync) so a failure
+            // in any of them still lands the task in Failed instead of stuck Running.
+            if (!alreadyClaimed)
+            {
+                var startResult = await _state.StartRunningAsync(taskId, now, ct);
+                if (!startResult.Ok)
+                {
+                    _logger.LogWarning("Task {TaskId} skipped: StartRunningAsync rejected ({Reason})", taskId, startResult.Reason);
+                    return;
+                }
+            }
+            else
+            {
+                // Queue-claimed dispatches skip StartRunningAsync, so broadcast TaskUpdated here
+                // (see RunAsync for the full rationale).
+                await _broadcaster.TaskUpdated(taskId);
+            }
+            await _broadcaster.TaskStarted(slot, taskId, now);
+
+            await _skillSeeder.SeedAsync(runDir, resolvedConfig.SkillNames, wtCtx is not null, ct);
+
+            (mcpToken, mcpConfigPath, resolvedConfig) = await SetupMcpConfigAsync(task, resolvedConfig, ct, token => mcpToken = token);
+
             var nextRunNumber = lastRun.RunNumber + 1;
             var result = await RunOnceAsync(taskId, task.Number, task.Title, slot, runDir, resolvedConfig, nextRunNumber, false, followUpPrompt, ct);
 
@@ -290,8 +293,12 @@ public sealed class TaskRunner
         }
         finally
         {
-            _tokens.Unregister(mcpToken);
+            if (mcpToken is not null)
-            try { File.Delete(mcpConfigPath); } catch { /* best effort */ }
+            {
+                _tokens.Unregister(mcpToken);
+                if (mcpConfigPath is not null)
+                    try { File.Delete(mcpConfigPath); } catch { /* best effort */ }
+            }
         }
     }
 
@@ -300,11 +307,15 @@ public sealed class TaskRunner
     // SuggestImprovement for filing out-of-scope follow-ups. Used by both a fresh run and a
     // --resume continuation, which needs the same wiring or the resumed session loses every
     // mcp__claudedo_run__* tool.
+    // onTokenRegistered fires the instant the token is registered, before any I/O that could
+    // still fail — callers assign it straight into their outer (pre-try) mcpToken variable so a
+    // later throw in this method still leaves the finally block able to unregister it.
     private async Task<(string McpToken, string McpConfigPath, ClaudeRunConfig Config)> SetupMcpConfigAsync(
-        TaskEntity task, ClaudeRunConfig resolvedConfig, CancellationToken ct)
+        TaskEntity task, ClaudeRunConfig resolvedConfig, CancellationToken ct, Action<string>? onTokenRegistered = null)
     {
         var mcpToken = TaskRunTokenRegistry.GenerateToken();
         _tokens.Register(mcpToken, task.Id);
+        onTokenRegistered?.Invoke(mcpToken);
         Directory.CreateDirectory(_cfg.LogRoot);
         var mcpConfigPath = Path.Combine(_cfg.LogRoot, $"{task.Id}_mcp.json");
         await File.WriteAllTextAsync(mcpConfigPath, BuildRunMcpConfigJson(mcpToken), ct);

tests/ClaudeDo.Worker.Tests/Infrastructure/FakeSessionSkillSeeder.cs

@@ -9,10 +9,15 @@ public sealed class FakeSessionSkillSeeder : ISessionSkillSeeder
     public readonly record struct Call(string WorkingDir, IReadOnlyList<string> SkillNames, bool IsWorktree);
     public readonly ConcurrentQueue<Call> Calls = new();
 
+    // Lets a test simulate a seeding failure (e.g. a broken skill install) to verify callers
+    // don't leave a task stuck Running when SeedAsync throws.
+    public Exception? ThrowOnSeed;
+
     public Task SeedAsync(string workingDir, IReadOnlyList<string> skillNames, bool isWorktree, CancellationToken ct)
     {
         Interlocked.Increment(ref CallCount);
         Calls.Enqueue(new Call(workingDir, skillNames, isWorktree));
+        if (ThrowOnSeed is not null) throw ThrowOnSeed;
         return Task.CompletedTask;
     }
 }

tests/ClaudeDo.Worker.Tests/Runner/ContinueAsyncExceptionTests.cs

@@ -4,6 +4,7 @@ using ClaudeDo.Data.Repositories;
 using ClaudeDo.Worker.Config;
 using ClaudeDo.Worker.Hub;
 using ClaudeDo.Worker.Runner;
+using ClaudeDo.Worker.Skills;
 using ClaudeDo.Worker.Tests.Infrastructure;
 using Microsoft.Extensions.Logging.Abstractions;
 using TaskStatus = ClaudeDo.Data.Models.TaskStatus;
@@ -29,14 +30,101 @@ public sealed class ContinueAsyncExceptionTests : IDisposable
 
     public void Dispose() { _db.Dispose(); try { Directory.Delete(_tempDir, true); } catch { } }
 
-    private TaskRunner BuildRunner(IClaudeProcess claude, ClaudeDoDbContext ctx)
+    private TaskRunner BuildRunner(
+        IClaudeProcess claude, ClaudeDoDbContext ctx,
+        ISessionSkillSeeder? skillSeeder = null, TaskRunTokenRegistry? tokens = null, WorkerConfig? cfg = null)
     {
         var dbFactory = _db.CreateFactory();
         var broadcaster = new HubBroadcaster(new CapturingHubContext());
         var state = TaskStateServiceBuilder.Build(dbFactory).State;
-        var wt = new WorktreeManager(new ClaudeDo.Data.Git.GitService(), dbFactory, _cfg, NullLogger<WorktreeManager>.Instance);
+        var effectiveCfg = cfg ?? _cfg;
-        return new TaskRunner(claude, dbFactory, broadcaster, wt, new ClaudeArgsBuilder(), _cfg,
+        var wt = new WorktreeManager(new ClaudeDo.Data.Git.GitService(), dbFactory, effectiveCfg, NullLogger<WorktreeManager>.Instance);
-            NullLogger<TaskRunner>.Instance, state, new TaskRunTokenRegistry(), new AttachmentStore(), new FakeSessionSkillSeeder(), new FakeTranscriptUsageReader());
+        return new TaskRunner(claude, dbFactory, broadcaster, wt, new ClaudeArgsBuilder(), effectiveCfg,
+            NullLogger<TaskRunner>.Instance, state, tokens ?? new TaskRunTokenRegistry(), new AttachmentStore(),
+            skillSeeder ?? new FakeSessionSkillSeeder(), new FakeTranscriptUsageReader());
+    }
+
+    private async Task<string> SeedContinuableTaskAsync(string sessionId)
+    {
+        string listId, taskId;
+        using var ctx = _db.CreateContext();
+        listId = Guid.NewGuid().ToString();
+        ctx.Lists.Add(new ListEntity { Id = listId, Name = "L", WorkingDir = null, CreatedAt = DateTime.UtcNow });
+
+        taskId = Guid.NewGuid().ToString();
+        ctx.Tasks.Add(new TaskEntity
+        {
+            Id = taskId,
+            ListId = listId,
+            Title = "Continue me",
+            Status = TaskStatus.WaitingForReview,
+            CreatedAt = DateTime.UtcNow,
+        });
+        await ctx.SaveChangesAsync();
+
+        await new TaskRunRepository(ctx).AddAsync(new TaskRunEntity
+        {
+            Id = Guid.NewGuid().ToString(),
+            TaskId = taskId,
+            RunNumber = 1,
+            IsRetry = false,
+            Prompt = "original prompt",
+            SessionId = sessionId,
+            StartedAt = DateTime.UtcNow.AddMinutes(-5),
+            FinishedAt = DateTime.UtcNow.AddMinutes(-1),
+            ExitCode = 0,
+            ResultMarkdown = "first result",
+        });
+        return taskId;
+    }
+
+    // Bug: SeedAsync used to run after the Running claim but before ContinueAsync's own
+    // try/catch started, so a throw here propagated straight out of ContinueAsync to the
+    // caller (OverrideSlotService.RunContinueInSlotAsync), which only logs — leaving the task
+    // stuck Running forever.
+    [Fact]
+    public async Task ContinueAsync_SkillSeederThrowsAfterClaim_MarksTaskFailed_NotStuckRunning()
+    {
+        var taskId = await SeedContinuableTaskAsync("sess-seed-throws");
+        var throwingSeeder = new FakeSessionSkillSeeder { ThrowOnSeed = new InvalidOperationException("skill seed exploded") };
+
+        using var ctx2 = _db.CreateContext();
+        var runner = BuildRunner(new FakeClaudeProcess(), ctx2, skillSeeder: throwingSeeder);
+
+        await runner.ContinueAsync(taskId, "please continue", "slot-1", CancellationToken.None);
+
+        using var verify = _db.CreateContext();
+        var task = await new TaskRepository(verify).GetByIdAsync(taskId);
+        Assert.NotNull(task);
+        Assert.Equal(TaskStatus.Failed, task.Status);
+    }
+
+    // Bug: SetupMcpConfigAsync registers the per-run MCP token *before* it can fail (writing
+    // the config file to disk). When it threw after Register but before the try/catch existed,
+    // the token was never unregistered — a leak in TaskRunTokenRegistry on top of the stuck
+    // Running task.
+    [Fact]
+    public async Task ContinueAsync_McpSetupThrowsAfterTokenRegistered_MarksTaskFailed_AndUnregistersToken()
+    {
+        var taskId = await SeedContinuableTaskAsync("sess-mcp-throws");
+
+        // SetupMcpConfigAsync calls Directory.CreateDirectory(cfg.LogRoot) right after
+        // registering the token; pointing LogRoot at an existing file makes that call throw.
+        var badLogRoot = Path.Combine(_tempDir, "logroot_is_a_file");
+        await File.WriteAllTextAsync(badLogRoot, "not a directory");
+        var cfg = new WorkerConfig { SandboxRoot = _tempDir, LogRoot = badLogRoot };
+        var tokens = new TaskRunTokenRegistry();
+
+        using var ctx2 = _db.CreateContext();
+        var runner = BuildRunner(new FakeClaudeProcess(), ctx2, tokens: tokens, cfg: cfg);
+
+        await runner.ContinueAsync(taskId, "please continue", "slot-1", CancellationToken.None);
+
+        using var verify = _db.CreateContext();
+        var task = await new TaskRepository(verify).GetByIdAsync(taskId);
+        Assert.NotNull(task);
+        Assert.Equal(TaskStatus.Failed, task.Status);
+        Assert.False(tokens.HasAnyTokenFor(taskId));
     }
 
     [Fact]

tests/ClaudeDo.Worker.Tests/Services/QueueServiceSlotFailureTests.cs

@@ -9,6 +9,7 @@ using ClaudeDo.Worker.Runner;
 using ClaudeDo.Worker.Tests.Infrastructure;
 using ClaudeDo.Worker.Usage;
 using Microsoft.Data.Sqlite;
+using Microsoft.EntityFrameworkCore;
 using Microsoft.Extensions.Logging.Abstractions;
 using TaskStatus = ClaudeDo.Data.Models.TaskStatus;
 
@@ -180,7 +181,7 @@ public sealed class QueueServiceSlotFailureTests : IDisposable
     }
 
     [Fact]
-    public async Task A_cancelled_slot_run_does_not_mark_the_task_Failed()
+    public async Task A_cancelled_slot_run_is_marked_Cancelled_not_left_Running()
     {
         var listId = await SeedListAsync();
         var taskId = Guid.NewGuid().ToString();
@@ -203,21 +204,105 @@ public sealed class QueueServiceSlotFailureTests : IDisposable
         await service.StartAsync(outerCts.Token);
         waker.Wake();
 
-        // Wait for the slot to be claimed and then released again (RunInSlotAsync's
+        var reloaded = await PollUntilLeftQueuedAsync(taskId);
-        // ContinueWith removes it once the catch block — ours or a stray one — finishes).
-        var deadline = DateTime.UtcNow.AddSeconds(10);
-        while (service.GetActive().Any(a => a.taskId == taskId) && DateTime.UtcNow < deadline)
-            await Task.Delay(25);
-        await Task.Delay(100); // let the fire-and-forget continuation fully settle
 
-        TaskEntity? reloaded;
+        // This reproduces a plain Stop (QueueService.CancelTask cancels the slot's CTS
-        using (var verify = _db.CreateContext())
+        // directly, without going through TaskStateService) landing during the pre-dispatch DB
-            reloaded = await new TaskRepository(verify).GetByIdAsync(taskId);
+        // reads — before TaskRunner's own claim/try-catch ever starts. Nothing else writes a
-
+        // terminal status for that window, so the OCE catch itself must close it.
-        // The picker's atomic claim already flipped it to Running; the cancelled slot run must
+        Assert.Equal(TaskStatus.Cancelled, reloaded!.Status);
-        // leave it there rather than flipping it to Failed.
+        Assert.Contains(hub.Proxy.Calls,
-        Assert.Equal(TaskStatus.Running, reloaded!.Status);
-        Assert.DoesNotContain(hub.Proxy.Calls,
             c => c.Method == "TaskUpdated" && (string)c.Args[0]! == taskId);
     }
+
+    // A_cancelled_slot_run_is_marked_Cancelled_not_left_Running above reproduces a plain Stop.
+    // TaskStateService.CancelAsync (the hub's CancelReview action) is a different origin: it
+    // writes Cancelled *before* it cancels the run's CTS via RunCancellationRegistry, so by the
+    // time RunInSlotAsync's OCE catch runs, the task has already left Running. The catch must
+    // recognize that and not blindly stomp whatever terminal status is already there.
+    [Fact]
+    public async Task A_cancelled_slot_run_does_not_stomp_a_status_already_written_terminal()
+    {
+        var listId = await SeedListAsync();
+        var taskId = Guid.NewGuid().ToString();
+
+        using (var ctx = _db.CreateContext())
+        {
+            ctx.Tasks.Add(new TaskEntity
+            {
+                Id = taskId, ListId = listId, Title = "T", Status = TaskStatus.Queued,
+                CreatedAt = DateTime.UtcNow,
+            });
+            await ctx.SaveChangesAsync();
+        }
+
+        var outerCts = new CancellationTokenSource();
+        var realPicker = new QueuePicker(_db.CreateFactory());
+        var picker = new ClaimMarkFailedThenCancelPicker(realPicker, outerCts, _db);
+        var (service, hub, waker) = CreateService(picker);
+
+        await service.StartAsync(outerCts.Token);
+        waker.Wake();
+
+        var reloaded = await PollUntilLeftQueuedAsync(taskId);
+
+        // Some other terminal write (simulating CancelAsync/FailAsync having already run) must
+        // survive — the OCE catch must not overwrite it with Cancelled just because it observed
+        // a cancellation.
+        Assert.Equal(TaskStatus.Failed, reloaded!.Status);
+    }
+
+    // Polls until the task leaves its initial Queued status (the claim + dispatch + OCE-catch
+    // chain settling), or gives up at the deadline; then waits a further grace period so the
+    // fire-and-forget continuation (queue-slot cleanup, any terminal-status write) fully lands
+    // before the caller reads the final state.
+    private async Task<TaskEntity?> PollUntilLeftQueuedAsync(string taskId)
+    {
+        TaskEntity? reloaded = null;
+        var deadline = DateTime.UtcNow.AddSeconds(10);
+        while (DateTime.UtcNow < deadline)
+        {
+            using var verify = _db.CreateContext();
+            reloaded = await new TaskRepository(verify).GetByIdAsync(taskId);
+            if (reloaded!.Status != TaskStatus.Queued) break;
+            await Task.Delay(25);
+        }
+        await Task.Delay(150); // let the fire-and-forget continuation fully settle
+
+        using var final = _db.CreateContext();
+        return await new TaskRepository(final).GetByIdAsync(taskId);
+    }
+
+    // Simulates a terminal status already having been written (by TaskStateService, from some
+    // other origin) between the picker's claim and RunInSlotAsync's OCE catch running.
+    private sealed class ClaimMarkFailedThenCancelPicker : IQueuePicker
+    {
+        private readonly IQueuePicker _inner;
+        private readonly CancellationTokenSource _cancelAfterClaim;
+        private readonly DbFixture _db;
+
+        public ClaimMarkFailedThenCancelPicker(IQueuePicker inner, CancellationTokenSource cancelAfterClaim, DbFixture db)
+        {
+            _inner = inner;
+            _cancelAfterClaim = cancelAfterClaim;
+            _db = db;
+        }
+
+        public async Task<TaskEntity?> ClaimNextAsync(DateTime now, CancellationToken ct)
+        {
+            var claimed = await _inner.ClaimNextAsync(now, ct);
+            if (claimed is not null)
+            {
+                using (var ctx = _db.CreateContext())
+                {
+                    await ctx.Tasks.Where(t => t.Id == claimed.Id)
+                        .ExecuteUpdateAsync(s => s
+                            .SetProperty(t => t.Status, TaskStatus.Failed)
+                            .SetProperty(t => t.FinishedAt, DateTime.UtcNow));
+                }
+                _cancelAfterClaim.Cancel();
+            }
+            return claimed;
+        }
+    }
 }