From 06d72a121564b5adcabcc9efae9ca68eeb3129d2 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Carlos=20Flor=C3=AAncio?= <carlosmflorencio@gmail.com>
Date: Mon, 27 Jul 2026 16:14:13 +0100
Subject: [PATCH] fix: enforce workflow-step WIP limits on task creation
 (#1980)

---
 .../backend/internal/mcp/handlers/handlers.go |  25 ++-
 .../internal/mcp/handlers/handlers_test.go    |   7 +
 .../orchestrator/event_handlers_github.go     |  37 +++-
 .../event_handlers_github_review_test.go      |  90 ++++++++
 .../orchestrator/event_handlers_workflow.go   |  63 +++---
 .../event_handlers_workflow_test.go           |  88 ++++++++
 .../internal/orchestrator/watcher_dispatch.go |  11 +-
 .../orchestrator/watcher_dispatch_test.go     |  21 ++
 .../internal/orchestrator/workflow_store.go   |  22 --
 apps/backend/internal/task/handlers/errors.go |   4 +
 .../internal/task/handlers/errors_test.go     |  15 ++
 .../task/handlers/task_http_handlers_test.go  |  36 ++++
 .../task/handlers/task_ws_handlers.go         |   3 +
 .../internal/task/repository/interface.go     |   2 +
 .../internal/task/repository/sqlite/task.go   |  56 ++++-
 .../internal/task/repository/task_wip_test.go |  99 +++++++++
 apps/backend/internal/task/service/service.go |   1 +
 .../internal/task/service/service_tasks.go    |  30 ++-
 .../task/service/service_tasks_wip_test.go    | 121 +++++++++++
 .../internal/task/service/service_workflow.go |   5 +-
 .../internal/workflow/models/errors.go        |  30 +++
 .../e2e/tests/kanban/mobile-kanban.spec.ts    |   4 +-
 docs/plans/wip-limited-task-creation/plan.md  | 199 ++++++++++++++++++
 .../task-01-atomic-repository-admission.md    |  65 ++++++
 .../task-02-task-service-enforcement.md       |  63 ++++++
 ...-conflict-adapters-and-watcher-deferral.md |  78 +++++++
 .../task-04-public-documentation.md           |  56 +++++
 docs/public/tasks-and-workflows.md            |  19 ++
 docs/public/workflow-tips.md                  |  15 ++
 docs/specs/tasks/wip-limit-pull-system.md     |  61 +++++-
 30 files changed, 1239 insertions(+), 87 deletions(-)
 create mode 100644 apps/backend/internal/task/repository/task_wip_test.go
 create mode 100644 apps/backend/internal/task/service/service_tasks_wip_test.go
 create mode 100644 apps/backend/internal/workflow/models/errors.go
 create mode 100644 docs/plans/wip-limited-task-creation/plan.md
 create mode 100644 docs/plans/wip-limited-task-creation/task-01-atomic-repository-admission.md
 create mode 100644 docs/plans/wip-limited-task-creation/task-02-task-service-enforcement.md
 create mode 100644 docs/plans/wip-limited-task-creation/task-03-conflict-adapters-and-watcher-deferral.md
 create mode 100644 docs/plans/wip-limited-task-creation/task-04-public-documentation.md

diff --git a/apps/backend/internal/mcp/handlers/handlers.go b/apps/backend/internal/mcp/handlers/handlers.go
index 2c72dd833a..2894e83a20 100644
--- a/apps/backend/internal/mcp/handlers/handlers.go
+++ b/apps/backend/internal/mcp/handlers/handlers.go
@@ -651,14 +651,12 @@ func (h *Handlers) handleCreateTask(ctx context.Context, msg *ws.Message) (*ws.M
 	})
 	if err != nil {
 		h.logger.Error("failed to create task", zap.Error(err))
-		// Defense-in-depth: resolveTaskRepositories already catches this for the
-		// MCP path, but non-MCP callers (UI, internal engine) reach here directly.
-		if errors.Is(err, service.ErrSubtaskDepthExceeded) ||
-			errors.Is(err, service.ErrInvalidTaskWorkflow) ||
-			isMCPWorkflowNotFoundError(err) {
-			return ws.NewError(msg.ID, msg.Action, ws.ErrorCodeValidation, err.Error(), nil)
+		code := classifyCreateTaskError(err)
+		message := "Failed to create task"
+		if code != ws.ErrorCodeInternalError {
+			message = err.Error()
 		}
-		return ws.NewError(msg.ID, msg.Action, ws.ErrorCodeInternalError, "Failed to create task", nil)
+		return ws.NewError(msg.ID, msg.Action, code, message, nil)
 	}
 
 	if h.handoffSvc != nil && workspacePolicy.NeedsAttachment() {
@@ -681,6 +679,19 @@ func (h *Handlers) handleCreateTask(ctx context.Context, msg *ws.Message) (*ws.M
 	return ws.NewResponse(msg.ID, msg.Action, dto.FromTask(task))
 }
 
+func classifyCreateTaskError(err error) string {
+	switch {
+	case errors.Is(err, service.ErrWIPLimitExceeded):
+		return ws.ErrorCodeConflict
+	case errors.Is(err, service.ErrSubtaskDepthExceeded),
+		errors.Is(err, service.ErrInvalidTaskWorkflow),
+		isMCPWorkflowNotFoundError(err):
+		return ws.ErrorCodeValidation
+	default:
+		return ws.ErrorCodeInternalError
+	}
+}
+
 // taskRepoResult holds the output of resolveTaskRepositories.
 type taskRepoResult struct {
 	Repos       []service.TaskRepositoryInput
diff --git a/apps/backend/internal/mcp/handlers/handlers_test.go b/apps/backend/internal/mcp/handlers/handlers_test.go
index c789d26acf..a6d49e94b7 100644
--- a/apps/backend/internal/mcp/handlers/handlers_test.go
+++ b/apps/backend/internal/mcp/handlers/handlers_test.go
@@ -196,6 +196,13 @@ func TestClassifyAddBranchError_UnresolvedBaseBranchIsValidation(t *testing.T) {
 	}
 }
 
+func TestClassifyCreateTaskErrorMapsWIPLimitToConflict(t *testing.T) {
+	err := fmt.Errorf("create task failed: %w", workflowmodels.NewWIPLimitError("review", 2, 2))
+	if got := classifyCreateTaskError(err); got != ws.ErrorCodeConflict {
+		t.Fatalf("expected ErrorCodeConflict, got %q", got)
+	}
+}
+
 // TestHandleAddBranchToTask_RejectsMultipleLocators pins the mutual-exclusion
 // check at the WS handler tier: supplying two of repository_id /
 // repository_url / local_path is an agent mistake that previously got
diff --git a/apps/backend/internal/orchestrator/event_handlers_github.go b/apps/backend/internal/orchestrator/event_handlers_github.go
index 19d84ca34e..8c56dc2b1b 100644
--- a/apps/backend/internal/orchestrator/event_handlers_github.go
+++ b/apps/backend/internal/orchestrator/event_handlers_github.go
@@ -2,6 +2,7 @@ package orchestrator
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"strconv"
 	"strings"
@@ -19,6 +20,10 @@ import (
 	wfmodels "github.com/kandev/kandev/internal/workflow/models"
 )
 
+func isWatcherCapacityRejection(err error) bool {
+	return errors.Is(err, wfmodels.ErrWIPLimitExceeded)
+}
+
 const (
 	issueWatchIDKey             = "issue_watch_id"
 	issueNumberKey              = "issue_number"
@@ -329,10 +334,18 @@ func (s *Service) createReviewTask(ctx context.Context, evt *github.NewReviewPRE
 	repositories := s.resolveReviewRepository(ctx, evt.WorkspaceID, pr)
 	task, err := s.reviewTaskCreator.CreateReviewTask(ctx, buildReviewTaskRequest(evt, repositories, repoSlug))
 	if err != nil {
-		s.logger.Error("failed to create review task",
-			zap.String("review_watch_id", evt.ReviewWatchID),
-			zap.Int("pr_number", pr.Number),
-			zap.Error(err))
+		if isWatcherCapacityRejection(err) {
+			s.logger.Info("review workflow step at capacity; deferring review task",
+				zap.String("review_watch_id", evt.ReviewWatchID),
+				zap.Int("pr_number", pr.Number),
+				zap.String("workflow_step_id", evt.WorkflowStepID),
+				zap.Error(err))
+		} else {
+			s.logger.Error("failed to create review task",
+				zap.String("review_watch_id", evt.ReviewWatchID),
+				zap.Int("pr_number", pr.Number),
+				zap.Error(err))
+		}
 		s.releaseReviewPR(ctx, evt)
 		return
 	}
@@ -1343,10 +1356,18 @@ func (s *Service) createIssueTask(ctx context.Context, evt *github.NewIssueEvent
 
 	task, err := s.issueTaskCreator.CreateIssueTask(ctx, req)
 	if err != nil {
-		s.logger.Error("failed to create issue task",
-			zap.String(issueWatchIDKey, evt.IssueWatchID),
-			zap.Int(issueNumberKey, issue.Number),
-			zap.Error(err))
+		if isWatcherCapacityRejection(err) {
+			s.logger.Info("issue workflow step at capacity; deferring issue task",
+				zap.String(issueWatchIDKey, evt.IssueWatchID),
+				zap.Int(issueNumberKey, issue.Number),
+				zap.String("workflow_step_id", evt.WorkflowStepID),
+				zap.Error(err))
+		} else {
+			s.logger.Error("failed to create issue task",
+				zap.String(issueWatchIDKey, evt.IssueWatchID),
+				zap.Int(issueNumberKey, issue.Number),
+				zap.Error(err))
+		}
 		s.releaseIssueWatch(ctx, evt)
 		return
 	}
diff --git a/apps/backend/internal/orchestrator/event_handlers_github_review_test.go b/apps/backend/internal/orchestrator/event_handlers_github_review_test.go
index 3de9ccd88f..9a16c5a494 100644
--- a/apps/backend/internal/orchestrator/event_handlers_github_review_test.go
+++ b/apps/backend/internal/orchestrator/event_handlers_github_review_test.go
@@ -5,6 +5,7 @@ import (
 	"testing"
 
 	"github.com/kandev/kandev/internal/github"
+	"github.com/kandev/kandev/internal/orchestrator/watcher"
 	"github.com/kandev/kandev/internal/task/models"
 	wfmodels "github.com/kandev/kandev/internal/workflow/models"
 )
@@ -13,11 +14,17 @@ import (
 type countingReviewTaskCreator struct {
 	calls  int
 	err    error
+	errs   []error
 	taskID string
 }
 
 func (c *countingReviewTaskCreator) CreateReviewTask(_ context.Context, _ *ReviewTaskRequest) (*models.Task, error) {
 	c.calls++
+	if len(c.errs) > 0 {
+		err := c.errs[0]
+		c.errs = c.errs[1:]
+		return nil, err
+	}
 	if c.err != nil {
 		return nil, c.err
 	}
@@ -137,3 +144,86 @@ func TestCreateReviewTask_ReleasesOnTaskCreationFailure(t *testing.T) {
 		t.Errorf("expected no Assign call after failure, got %d", ghSvc.assignCalls)
 	}
 }
+
+func TestGitHubWatcherTaskCreateError_ClassifiesWIPAsDeferred(t *testing.T) {
+	err := wfmodels.NewWIPLimitError("step1", 2, 2)
+	if !isWatcherCapacityRejection(err) {
+		t.Fatal("expected WIP-limit rejection to be classified as deferred capacity")
+	}
+}
+
+func TestCreateReviewTask_WIPRejectionIsRetriedOnLaterPoll(t *testing.T) {
+	svc, _ := setupReviewTaskTest(t)
+	ghSvc := &mockGitHubService{reserveReturn: true}
+	svc.SetGitHubService(ghSvc)
+	creator := &countingReviewTaskCreator{
+		errs:   []error{wfmodels.NewWIPLimitError("step1", 2, 2)},
+		taskID: "task-after-capacity",
+	}
+	svc.SetReviewTaskCreator(creator)
+
+	svc.createReviewTask(context.Background(), newReviewEvent())
+	svc.createReviewTask(context.Background(), newReviewEvent())
+
+	if creator.calls != 2 {
+		t.Fatalf("expected one rejected attempt and one retry, got %d calls", creator.calls)
+	}
+	if ghSvc.releaseCalls != 1 {
+		t.Fatalf("expected only the rejected attempt to release its reservation, got %d", ghSvc.releaseCalls)
+	}
+	if ghSvc.assignCalls != 1 || ghSvc.assignedTaskID != "task-after-capacity" {
+		t.Fatalf("expected retry to assign task ID, calls=%d task=%q", ghSvc.assignCalls, ghSvc.assignedTaskID)
+	}
+}
+
+func TestReviewWatchLifecycle_BootReadyStaysOnReviewUntilTurnComplete(t *testing.T) {
+	ctx := context.Background()
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "review-task", "review-session", "review")
+
+	steps := newMockStepGetter()
+	steps.steps["review"] = &wfmodels.WorkflowStep{
+		ID: "review", WorkflowID: "wf1", Name: "Review", Position: 0,
+		Events: wfmodels.StepEvents{
+			OnEnter:        []wfmodels.OnEnterAction{{Type: wfmodels.OnEnterAutoStartAgent}},
+			OnTurnComplete: []wfmodels.OnTurnCompleteAction{{Type: wfmodels.OnTurnCompleteMoveToNext}},
+		},
+	}
+	steps.steps["done"] = &wfmodels.WorkflowStep{ID: "done", WorkflowID: "wf1", Name: "Done", Position: 1}
+	agentMgr := &mockAgentManager{repoForExecutionLookup: repo}
+	svc := createTestServiceWithAgent(repo, steps, newMockTaskRepo(), agentMgr)
+
+	if !svc.shouldAutoStartStep(ctx, "review") {
+		t.Fatal("review step should auto-start an admitted watcher task")
+	}
+	svc.handleAgentBootReady(ctx, watcher.AgentEventData{TaskID: "review-task", SessionID: "review-session"})
+	bootTask, err := repo.GetTask(ctx, "review-task")
+	if err != nil {
+		t.Fatalf("load task after boot-ready: %v", err)
+	}
+	if bootTask.WorkflowStepID != "review" {
+		t.Fatalf("boot-ready must not advance review task, got step %q", bootTask.WorkflowStepID)
+	}
+
+	if err := repo.UpdateTaskSessionState(ctx, "review-session", models.TaskSessionStateRunning, ""); err != nil {
+		t.Fatalf("mark review turn running: %v", err)
+	}
+	task, err := repo.GetTask(ctx, "review-task")
+	if err != nil {
+		t.Fatalf("reload task before turn completion: %v", err)
+	}
+	session, err := repo.GetTaskSession(ctx, "review-session")
+	if err != nil {
+		t.Fatalf("reload session before turn completion: %v", err)
+	}
+	if !svc.processOnTurnComplete(ctx, task, session) {
+		t.Fatal("first genuine turn completion should advance the review task")
+	}
+	completedTask, err := repo.GetTask(ctx, "review-task")
+	if err != nil {
+		t.Fatalf("load task after turn completion: %v", err)
+	}
+	if completedTask.WorkflowStepID != "done" {
+		t.Fatalf("expected turn completion to move task to done, got %q", completedTask.WorkflowStepID)
+	}
+}
diff --git a/apps/backend/internal/orchestrator/event_handlers_workflow.go b/apps/backend/internal/orchestrator/event_handlers_workflow.go
index ba78565909..e082ee6462 100644
--- a/apps/backend/internal/orchestrator/event_handlers_workflow.go
+++ b/apps/backend/internal/orchestrator/event_handlers_workflow.go
@@ -255,30 +255,12 @@ func (s *Service) executeStepTransition(ctx context.Context, taskID, sessionID s
 		s.setSessionWaitingForInput(ctx, taskID, sessionID)
 		return
 	}
-	if err := s.validateTransitionWIPLimit(ctx, task, targetStep); err != nil {
-		s.logger.Warn("workflow transition rejected by WIP limit",
-			zap.String("task_id", taskID),
-			zap.String("to_step_id", toStepID),
-			zap.Error(err))
-		s.setSessionWaitingForInput(ctx, taskID, sessionID)
-		return
-	}
-
-	// Process on_exit actions for the step we're leaving (before the step change).
-	// Freshly load the session since the caller may not have it (legacy path).
-	exitSession, exitErr := s.repo.GetTaskSession(ctx, sessionID)
-	if exitErr != nil {
-		s.logger.Warn("failed to load session for on_exit",
-			zap.String("session_id", sessionID), zap.Error(exitErr))
-	} else {
-		s.processOnExit(ctx, taskID, exitSession, fromStep)
-	}
-
-	// Update the task's workflow step
+	// Atomically admit the target step before exit side effects. A capacity
+	// rejection must leave both the task and its source-step session state intact.
 	task.WorkflowStepID = toStepID
 	task.UpdatedAt = time.Now().UTC()
-	if err := s.repo.UpdateTask(ctx, task); err != nil {
-		s.logger.Error("failed to move task to next workflow step",
+	if err := s.updateTransitionTaskWithCapacity(ctx, task, targetStep); err != nil {
+		s.logger.Warn("workflow transition rejected or failed",
 			zap.String("task_id", taskID),
 			zap.String("from_step", fromStep.Name),
 			zap.String("to_step", targetStep.Name),
@@ -287,6 +269,16 @@ func (s *Service) executeStepTransition(ctx context.Context, taskID, sessionID s
 		return
 	}
 
+	// Process on_exit only after the transition is durably admitted. Freshly
+	// load the session since the caller may not have it (legacy path).
+	exitSession, exitErr := s.repo.GetTaskSession(ctx, sessionID)
+	if exitErr != nil {
+		s.logger.Warn("failed to load session for on_exit",
+			zap.String("session_id", sessionID), zap.Error(exitErr))
+	} else {
+		s.processOnExit(ctx, taskID, exitSession, fromStep)
+	}
+
 	// Publish task updated event via the task service so the payload carries
 	// the full context (session counts, primary session, repositories).
 	s.publishTaskUpdated(ctx, task)
@@ -346,22 +338,25 @@ func (s *Service) executeStepTransition(ctx context.Context, taskID, sessionID s
 	}
 }
 
-func (s *Service) validateTransitionWIPLimit(ctx context.Context, task *models.Task, targetStep *wfmodels.WorkflowStep) error {
-	if targetStep == nil || targetStep.WIPLimit <= 0 || task.WorkflowStepID == targetStep.ID {
-		return nil
+func (s *Service) updateTransitionTaskWithCapacity(
+	ctx context.Context,
+	task *models.Task,
+	targetStep *wfmodels.WorkflowStep,
+) error {
+	if targetStep == nil || targetStep.WIPLimit <= 0 {
+		return s.repo.UpdateTask(ctx, task)
 	}
-	limitsRepo, ok := s.repo.(workflowMoveLimitsRepository)
+	limitedRepo, ok := s.repo.(workflowLimitedMoveRepository)
 	if !ok {
 		return fmt.Errorf("WIP limit cannot be checked for workflow step %s", targetStep.ID)
 	}
-	occupants, err := limitsRepo.CountTasksByWorkflowStepExcludingTask(ctx, targetStep.ID, task.ID)
-	if err != nil {
-		return fmt.Errorf("count target workflow step tasks: %w", err)
-	}
-	if occupants >= targetStep.WIPLimit {
-		return fmt.Errorf("WIP limit exceeded for workflow step %s: limit %d already occupied", targetStep.ID, targetStep.WIPLimit)
-	}
-	return nil
+	return limitedRepo.UpdateTaskIfWorkflowStepHasCapacity(
+		ctx,
+		task,
+		targetStep.ID,
+		task.ID,
+		targetStep.WIPLimit,
+	)
 }
 
 // handleTaskMoved handles manual task step changes (drag-and-drop, stepper "Move here").
diff --git a/apps/backend/internal/orchestrator/event_handlers_workflow_test.go b/apps/backend/internal/orchestrator/event_handlers_workflow_test.go
index b7bdb3c30a..90c3c4ef8d 100644
--- a/apps/backend/internal/orchestrator/event_handlers_workflow_test.go
+++ b/apps/backend/internal/orchestrator/event_handlers_workflow_test.go
@@ -2,6 +2,7 @@ package orchestrator
 
 import (
 	"context"
+	"errors"
 	"sync"
 	"testing"
 	"time"
@@ -9,6 +10,7 @@ import (
 	"github.com/kandev/kandev/internal/events"
 	"github.com/kandev/kandev/internal/events/bus"
 	"github.com/kandev/kandev/internal/task/models"
+	wfmodels "github.com/kandev/kandev/internal/workflow/models"
 )
 
 // mockEventBus captures published events for assertion.
@@ -17,6 +19,92 @@ type mockEventBus struct {
 	events []publishedEvent
 }
 
+func TestUpdateTransitionTaskWithCapacity_RejectsFullLimitedStep(t *testing.T) {
+	ctx := context.Background()
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	if err := repo.CreateTask(ctx, &models.Task{
+		ID:             "occupant",
+		WorkspaceID:    "ws1",
+		WorkflowID:     "wf1",
+		WorkflowStepID: "step2",
+		Title:          "Occupant",
+		State:          "TODO",
+		Priority:       "medium",
+	}); err != nil {
+		t.Fatalf("create occupant: %v", err)
+	}
+
+	svc := createTestService(repo, newMockStepGetter(), newMockTaskRepo())
+	task, err := repo.GetTask(ctx, "t1")
+	if err != nil {
+		t.Fatalf("load task: %v", err)
+	}
+	task.WorkflowStepID = "step2"
+	target := &wfmodels.WorkflowStep{ID: "step2", WorkflowID: "wf1", WIPLimit: 1}
+
+	err = svc.updateTransitionTaskWithCapacity(ctx, task, target)
+	if !errors.Is(err, wfmodels.ErrWIPLimitExceeded) {
+		t.Fatalf("error=%v, want typed WIP rejection", err)
+	}
+	stored, err := repo.GetTask(ctx, "t1")
+	if err != nil {
+		t.Fatalf("reload task: %v", err)
+	}
+	if stored.WorkflowStepID != "step1" {
+		t.Fatalf("task moved despite full target, got %q", stored.WorkflowStepID)
+	}
+}
+
+func TestExecuteStepTransition_FullTargetLeavesOnExitStateIntact(t *testing.T) {
+	ctx := context.Background()
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	if err := repo.SetSessionMetadataKey(ctx, "s1", "plan_mode", true); err != nil {
+		t.Fatalf("enable plan mode: %v", err)
+	}
+	if err := repo.CreateTask(ctx, &models.Task{
+		ID:             "occupant",
+		WorkspaceID:    "ws1",
+		WorkflowID:     "wf1",
+		WorkflowStepID: "step2",
+		Title:          "Occupant",
+		State:          "TODO",
+		Priority:       "medium",
+	}); err != nil {
+		t.Fatalf("create occupant: %v", err)
+	}
+
+	steps := newMockStepGetter()
+	fromStep := &wfmodels.WorkflowStep{
+		ID: "step1", WorkflowID: "wf1", Name: "Source",
+		Events: wfmodels.StepEvents{OnExit: []wfmodels.OnExitAction{{
+			Type: wfmodels.OnExitDisablePlanMode,
+		}}},
+	}
+	steps.steps["step2"] = &wfmodels.WorkflowStep{
+		ID: "step2", WorkflowID: "wf1", Name: "Limited", WIPLimit: 1,
+	}
+	svc := createTestService(repo, steps, newMockTaskRepo())
+
+	svc.executeStepTransition(ctx, "t1", "s1", fromStep, "step2", false)
+
+	task, err := repo.GetTask(ctx, "t1")
+	if err != nil {
+		t.Fatalf("load task: %v", err)
+	}
+	if task.WorkflowStepID != "step1" {
+		t.Fatalf("task moved despite full target, got %q", task.WorkflowStepID)
+	}
+	session, err := repo.GetTaskSession(ctx, "s1")
+	if err != nil {
+		t.Fatalf("load session: %v", err)
+	}
+	if enabled, _ := session.Metadata["plan_mode"].(bool); !enabled {
+		t.Fatal("capacity rejection must not run source step's on_exit actions")
+	}
+}
+
 type publishedEvent struct {
 	Subject string
 	Event   *bus.Event
diff --git a/apps/backend/internal/orchestrator/watcher_dispatch.go b/apps/backend/internal/orchestrator/watcher_dispatch.go
index a4e818ba34..70dc8debcc 100644
--- a/apps/backend/internal/orchestrator/watcher_dispatch.go
+++ b/apps/backend/internal/orchestrator/watcher_dispatch.go
@@ -2,11 +2,13 @@ package orchestrator
 
 import (
 	"context"
+	"errors"
 	"sync"
 
 	"go.uber.org/zap"
 
 	"github.com/kandev/kandev/internal/common/logger"
+	workflowmodels "github.com/kandev/kandev/internal/workflow/models"
 )
 
 // WatcherDispatchCoordinator owns the cross-integration pipeline that turns
@@ -357,8 +359,13 @@ func (c *WatcherDispatchCoordinator) Dispatch(ctx context.Context, src WatcherSo
 
 	task, err := c.getTaskCreator().CreateIssueTask(ctx, req)
 	if err != nil {
-		c.logger.Error("watcher dispatch: create issue task failed",
-			zap.String("source", src.Name()), zap.Error(err))
+		if errors.Is(err, workflowmodels.ErrWIPLimitExceeded) {
+			c.logger.Info("watcher dispatch: workflow step at capacity; deferring issue task",
+				zap.String("source", src.Name()), zap.String("workflow_step_id", req.WorkflowStepID), zap.Error(err))
+		} else {
+			c.logger.Error("watcher dispatch: create issue task failed",
+				zap.String("source", src.Name()), zap.Error(err))
+		}
 		src.Release(ctx, evt)
 		return
 	}
diff --git a/apps/backend/internal/orchestrator/watcher_dispatch_test.go b/apps/backend/internal/orchestrator/watcher_dispatch_test.go
index 57e1f46b08..6375127f65 100644
--- a/apps/backend/internal/orchestrator/watcher_dispatch_test.go
+++ b/apps/backend/internal/orchestrator/watcher_dispatch_test.go
@@ -9,6 +9,7 @@ import (
 
 	"github.com/kandev/kandev/internal/common/logger"
 	"github.com/kandev/kandev/internal/task/models"
+	workflowmodels "github.com/kandev/kandev/internal/workflow/models"
 )
 
 // nopLogger returns a *logger.Logger backed by zap.NewNop() for use in tests.
@@ -247,6 +248,26 @@ func TestCoordinator_Dispatch_CreateError_Releases(t *testing.T) {
 	}
 }
 
+func TestCoordinator_Dispatch_WIPLimitRejection_ReleasesWithoutErrorSemantics(t *testing.T) {
+	src := &fakeWatcherSource{
+		name:      "linear",
+		reserveOK: true,
+		buildReq:  &IssueTaskRequest{WorkspaceID: "ws-1", WorkflowStepID: "step-1"},
+	}
+	tc := &fakeTaskCreator{createErr: workflowmodels.NewWIPLimitError("step-1", 2, 2)}
+	starter := &fakeTaskStarter{}
+	c := newTestCoordinator(t, tc, true, starter)
+
+	c.Dispatch(context.Background(), src, "evt")
+
+	if src.recordedRelease != 1 {
+		t.Fatalf("expected release after WIP rejection, got %d", src.recordedRelease)
+	}
+	if src.recordedAttach != 0 || starter.called != 0 {
+		t.Fatal("expected WIP rejection to stop before attach and auto-start")
+	}
+}
+
 func TestCoordinator_Dispatch_NoAutoStart(t *testing.T) {
 	src := &fakeWatcherSource{
 		name:      "linear",
diff --git a/apps/backend/internal/orchestrator/workflow_store.go b/apps/backend/internal/orchestrator/workflow_store.go
index 35b3512c9f..9224961f48 100644
--- a/apps/backend/internal/orchestrator/workflow_store.go
+++ b/apps/backend/internal/orchestrator/workflow_store.go
@@ -127,10 +127,6 @@ func (s *workflowStore) ApplyTransition(ctx context.Context, taskID, sessionID,
 	if err != nil {
 		return fmt.Errorf("load target step for transition: %w", err)
 	}
-	if err := s.validateWIPLimit(ctx, task, targetStep); err != nil {
-		return err
-	}
-
 	// Keep WorkflowID in sync with the target step's owning workflow. Most
 	// callers transition within the same workflow (targetStep.WorkflowID ==
 	// task.WorkflowID already), but applyPendingMove uses this path for
@@ -180,24 +176,6 @@ func (s *workflowStore) updateTransitionTask(ctx context.Context, task *models.T
 	return limitedRepo.UpdateTaskIfWorkflowStepHasCapacity(ctx, task, targetStep.ID, task.ID, targetStep.WIPLimit)
 }
 
-func (s *workflowStore) validateWIPLimit(ctx context.Context, task *models.Task, targetStep *wfmodels.WorkflowStep) error {
-	if targetStep == nil || targetStep.WIPLimit <= 0 || task.WorkflowStepID == targetStep.ID {
-		return nil
-	}
-	limitsRepo, ok := s.repo.(workflowMoveLimitsRepository)
-	if !ok {
-		return fmt.Errorf("WIP limit cannot be checked for workflow step %s", targetStep.ID)
-	}
-	occupants, err := limitsRepo.CountTasksByWorkflowStepExcludingTask(ctx, targetStep.ID, task.ID)
-	if err != nil {
-		return fmt.Errorf("count target workflow step tasks: %w", err)
-	}
-	if occupants >= targetStep.WIPLimit {
-		return fmt.Errorf("WIP limit exceeded for workflow step %s: limit %d already occupied", targetStep.ID, targetStep.WIPLimit)
-	}
-	return nil
-}
-
 func (s *workflowStore) pullNextTaskOnVacate(ctx context.Context, vacatedStepID, excludeTaskID string) {
 	vacatedStep := s.pullEnabledStep(ctx, vacatedStepID)
 	if vacatedStep == nil {
diff --git a/apps/backend/internal/task/handlers/errors.go b/apps/backend/internal/task/handlers/errors.go
index b13e4add11..18ae1a8326 100644
--- a/apps/backend/internal/task/handlers/errors.go
+++ b/apps/backend/internal/task/handlers/errors.go
@@ -17,6 +17,10 @@ func handleNotFound(c *gin.Context, log *logger.Logger, err error, fallback stri
 		c.JSON(http.StatusNotFound, gin.H{"error": fallback})
 		return
 	}
+	if errors.Is(err, service.ErrWIPLimitExceeded) {
+		c.JSON(http.StatusConflict, gin.H{"error": err.Error()})
+		return
+	}
 	if isValidationError(err) {
 		c.JSON(http.StatusBadRequest, gin.H{"error": err.Error()})
 		return
diff --git a/apps/backend/internal/task/handlers/errors_test.go b/apps/backend/internal/task/handlers/errors_test.go
index 4a645c2f91..7207b4fa04 100644
--- a/apps/backend/internal/task/handlers/errors_test.go
+++ b/apps/backend/internal/task/handlers/errors_test.go
@@ -5,11 +5,15 @@ import (
 	"errors"
 	"fmt"
 	"net"
+	"net/http"
+	"net/http/httptest"
 	"testing"
 
+	"github.com/gin-gonic/gin"
 	"github.com/kandev/kandev/internal/agent/runtime/lifecycle"
 	taskrepo "github.com/kandev/kandev/internal/task/repository/sqlite"
 	"github.com/kandev/kandev/internal/task/service"
+	wfmodels "github.com/kandev/kandev/internal/workflow/models"
 )
 
 // TestErrorsAreClassifiable verifies that the package's error-classification
@@ -42,6 +46,16 @@ func TestErrorsAreClassifiable(t *testing.T) {
 			t.Errorf("untyped lookalike must no longer classify")
 		}
 	})
+
+	t.Run("WIP limit is a conflict for task creation", func(t *testing.T) {
+		gin.SetMode(gin.TestMode)
+		rec := httptest.NewRecorder()
+		ctx, _ := gin.CreateTestContext(rec)
+		handleNotFound(ctx, newTestLogger(t), fmt.Errorf("create task: %w", wfmodels.NewWIPLimitError("review", 2, 2)), "task not created")
+		if rec.Code != http.StatusConflict {
+			t.Fatalf("status=%d, want %d", rec.Code, http.StatusConflict)
+		}
+	})
 }
 
 // TestIsTimeoutError pins the UX-classification contract for the prompt
@@ -75,6 +89,7 @@ func TestIsTimeoutError(t *testing.T) {
 			}
 		})
 	}
+
 }
 
 type timeoutNetErr struct{}
diff --git a/apps/backend/internal/task/handlers/task_http_handlers_test.go b/apps/backend/internal/task/handlers/task_http_handlers_test.go
index 1c8c13ba7c..2af07aeaa5 100644
--- a/apps/backend/internal/task/handlers/task_http_handlers_test.go
+++ b/apps/backend/internal/task/handlers/task_http_handlers_test.go
@@ -320,6 +320,14 @@ type missingWorkflowCreateTaskRepo struct {
 	captureCreateTaskRepo
 }
 
+type wipCreateTaskRepo struct {
+	captureCreateTaskRepo
+}
+
+func (*wipCreateTaskRepo) CreateTask(_ context.Context, _ *models.Task) error {
+	return wfmodels.NewWIPLimitError("review", 2, 2)
+}
+
 func (*missingWorkflowCreateTaskRepo) GetWorkflow(_ context.Context, id string) (*models.Workflow, error) {
 	return nil, fmt.Errorf("workflow not found: %s", id)
 }
@@ -703,6 +711,34 @@ func TestWSCreateTaskReturnsValidationErrorForMissingWorkflow(t *testing.T) {
 	assert.Contains(t, payload.Message, "workflow not found")
 }
 
+func TestWSCreateTaskReturnsConflictForWIPLimit(t *testing.T) {
+	log := newTestLogger(t)
+	repo := &wipCreateTaskRepo{}
+	svc := service.NewService(service.Repos{
+		Workspaces: repo, Tasks: repo, TaskRepos: repo,
+		Workflows: repo, Messages: repo, Turns: repo,
+		Sessions: repo, GitSnapshots: repo, RepoEntities: repo,
+		Executors: repo, Environments: repo, TaskEnvironments: repo,
+		Reviews: repo,
+	}, nil, log, service.RepositoryDiscoveryConfig{})
+	h := &TaskHandlers{service: svc, logger: log}
+
+	msg, err := ws.NewRequest("msg-1", ws.ActionTaskCreate, map[string]any{
+		"workspace_id": "ws-1",
+		"workflow_id":  "wf-1",
+		"title":        "Full review step",
+	})
+	require.NoError(t, err)
+
+	resp, err := h.wsCreateTask(context.Background(), msg)
+	require.NoError(t, err)
+	require.Equal(t, ws.MessageTypeError, resp.Type)
+	var payload ws.ErrorPayload
+	require.NoError(t, json.Unmarshal(resp.Payload, &payload))
+	assert.Equal(t, ws.ErrorCodeConflict, payload.Code)
+	assert.Contains(t, payload.Message, "WIP limit")
+}
+
 func TestWSCreateTaskReturnsValidationErrorForStepOutsideWorkflow(t *testing.T) {
 	log := newTestLogger(t)
 	repo := &captureCreateTaskRepo{}
diff --git a/apps/backend/internal/task/handlers/task_ws_handlers.go b/apps/backend/internal/task/handlers/task_ws_handlers.go
index b289d0ff59..3ed2f49646 100644
--- a/apps/backend/internal/task/handlers/task_ws_handlers.go
+++ b/apps/backend/internal/task/handlers/task_ws_handlers.go
@@ -171,6 +171,9 @@ func (h *TaskHandlers) wsCreateTask(ctx context.Context, msg *ws.Message) (*ws.M
 	})
 	if err != nil {
 		h.logger.Error("failed to create task", zap.Error(err))
+		if errors.Is(err, service.ErrWIPLimitExceeded) {
+			return ws.NewError(msg.ID, msg.Action, ws.ErrorCodeConflict, err.Error(), nil)
+		}
 		if isTaskCreateValidationError(err) {
 			return ws.NewError(msg.ID, msg.Action, ws.ErrorCodeValidation, err.Error(), nil)
 		}
diff --git a/apps/backend/internal/task/repository/interface.go b/apps/backend/internal/task/repository/interface.go
index 4bf6bb6ca2..af61895fd0 100644
--- a/apps/backend/internal/task/repository/interface.go
+++ b/apps/backend/internal/task/repository/interface.go
@@ -7,6 +7,7 @@ import (
 	agentdto "github.com/kandev/kandev/internal/agent/dto"
 	"github.com/kandev/kandev/internal/task/models"
 	"github.com/kandev/kandev/internal/task/repository/repoerrors"
+	wfmodels "github.com/kandev/kandev/internal/workflow/models"
 	v1 "github.com/kandev/kandev/pkg/api/v1"
 )
 
@@ -15,6 +16,7 @@ var ErrWorkspaceNotFound = repoerrors.ErrWorkspaceNotFound
 var ErrTaskNotFound = repoerrors.ErrTaskNotFound
 var ErrTaskPlanNotFound = repoerrors.ErrTaskPlanNotFound
 var ErrRepositoryNotFound = repoerrors.ErrRepositoryNotFound
+var ErrWIPLimitExceeded = wfmodels.ErrWIPLimitExceeded
 
 // WorkspaceRepository handles workspace CRUD.
 type WorkspaceRepository interface {
diff --git a/apps/backend/internal/task/repository/sqlite/task.go b/apps/backend/internal/task/repository/sqlite/task.go
index 461299f020..56fbc11258 100644
--- a/apps/backend/internal/task/repository/sqlite/task.go
+++ b/apps/backend/internal/task/repository/sqlite/task.go
@@ -16,6 +16,7 @@ import (
 	"github.com/kandev/kandev/internal/db/dialect"
 	"github.com/kandev/kandev/internal/task/models"
 	usermodels "github.com/kandev/kandev/internal/user/models"
+	wfmodels "github.com/kandev/kandev/internal/workflow/models"
 	v1 "github.com/kandev/kandev/pkg/api/v1"
 )
 
@@ -139,6 +140,20 @@ func runnerProjection(alias string) string {
 // (ADR 0005 Wave F); when the request carries AssigneeAgentProfileID we
 // upsert a 'runner' row in workflow_step_participants instead.
 func (r *Repository) CreateTask(ctx context.Context, task *models.Task) error {
+	return r.createTask(ctx, task, "", 0)
+}
+
+// CreateTaskIfWorkflowStepHasCapacity atomically admits a task into a
+// WIP-limited workflow step. The occupancy check and insert share one writer
+// transaction, so concurrent watcher events cannot overfill the step.
+func (r *Repository) CreateTaskIfWorkflowStepHasCapacity(ctx context.Context, task *models.Task, targetStepID string, limit int) error {
+	if limit <= 0 {
+		return r.CreateTask(ctx, task)
+	}
+	return r.createTask(ctx, task, targetStepID, limit)
+}
+
+func (r *Repository) createTask(ctx context.Context, task *models.Task, targetStepID string, limit int) error {
 	if task.ID == "" {
 		task.ID = uuid.New().String()
 	}
@@ -163,6 +178,11 @@ func (r *Repository) CreateTask(ctx context.Context, task *models.Task) error {
 	if err != nil {
 		return err
 	}
+	defer func() { _ = tx.Rollback() }()
+
+	if err := r.ensureWorkflowStepCapacity(ctx, tx, targetStepID, limit); err != nil {
+		return err
+	}
 
 	_, err = tx.ExecContext(ctx, r.db.Rebind(`
 		INSERT INTO tasks (id, workspace_id, workflow_id, workflow_step_id, title, description, state, priority, position, metadata, is_ephemeral, parent_id, created_at, updated_at, origin, project_id, labels, identifier)
@@ -187,6 +207,36 @@ func (r *Repository) CreateTask(ctx context.Context, task *models.Task) error {
 	return tx.Commit()
 }
 
+func (r *Repository) ensureWorkflowStepCapacity(ctx context.Context, tx *sql.Tx, targetStepID string, limit int) error {
+	if limit <= 0 {
+		return nil
+	}
+	if err := lockWorkflowStepForCapacity(ctx, tx, r.db.DriverName(), r.db.Rebind, targetStepID); err != nil {
+		return err
+	}
+	var occupants int
+	if err := tx.QueryRowContext(ctx, r.db.Rebind(`
+		SELECT COUNT(*) FROM tasks
+		WHERE workflow_step_id = ?
+		  AND archived_at IS NULL
+		  AND is_ephemeral = 0
+	`), targetStepID).Scan(&occupants); err != nil {
+		return err
+	}
+	if occupants >= limit {
+		return wfmodels.NewWIPLimitError(targetStepID, limit, occupants)
+	}
+	return nil
+}
+
+func lockWorkflowStepForCapacity(ctx context.Context, tx *sql.Tx, driver string, rebind func(string) string, stepID string) error {
+	if !dialect.IsPostgres(driver) {
+		return nil
+	}
+	var lockedID string
+	return tx.QueryRowContext(ctx, rebind(`SELECT id FROM workflow_steps WHERE id = ? FOR UPDATE`), stepID).Scan(&lockedID)
+}
+
 // upsertRunnerInTx writes (or replaces) a 'runner' participant row for
 // (stepID, taskID) inside the provided transaction. Mirrors
 // workflow.Repository.SetTaskRunner but reuses the caller's tx.
@@ -367,6 +417,10 @@ func (r *Repository) UpdateTaskIfWorkflowStepHasCapacity(ctx context.Context, ta
 	}
 	defer func() { _ = tx.Rollback() }()
 
+	if err := lockWorkflowStepForCapacity(ctx, tx, r.db.DriverName(), r.db.Rebind, targetStepID); err != nil {
+		return err
+	}
+
 	var occupants int
 	if err := tx.QueryRowContext(ctx, r.db.Rebind(`
 		SELECT COUNT(*) FROM tasks
@@ -378,7 +432,7 @@ func (r *Repository) UpdateTaskIfWorkflowStepHasCapacity(ctx context.Context, ta
 		return err
 	}
 	if occupants >= limit {
-		return fmt.Errorf("WIP limit exceeded for workflow step %s: limit %d already occupied", targetStepID, limit)
+		return wfmodels.NewWIPLimitError(targetStepID, limit, occupants)
 	}
 
 	result, err := tx.ExecContext(ctx, r.db.Rebind(`
diff --git a/apps/backend/internal/task/repository/task_wip_test.go b/apps/backend/internal/task/repository/task_wip_test.go
new file mode 100644
index 0000000000..360c7ee396
--- /dev/null
+++ b/apps/backend/internal/task/repository/task_wip_test.go
@@ -0,0 +1,99 @@
+package repository
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"sync"
+	"testing"
+
+	"github.com/kandev/kandev/internal/task/models"
+	wfmodels "github.com/kandev/kandev/internal/workflow/models"
+	v1 "github.com/kandev/kandev/pkg/api/v1"
+)
+
+type workflowStepCapacityCreator interface {
+	CreateTaskIfWorkflowStepHasCapacity(context.Context, *models.Task, string, int) error
+}
+
+func TestUpdateTaskIfWorkflowStepHasCapacity_ReturnsTypedWIPError(t *testing.T) {
+	repo, cleanup := createTestSQLiteRepo(t)
+	defer cleanup()
+	ctx := context.Background()
+
+	if err := repo.CreateTask(ctx, &models.Task{
+		ID: "wip-existing", WorkspaceID: "wip-workspace", WorkflowID: "wip-workflow",
+		WorkflowStepID: "wip-step", Title: "Existing", State: v1.TaskStateCreated,
+	}); err != nil {
+		t.Fatalf("seed existing task: %v", err)
+	}
+	candidate := &models.Task{
+		ID: "wip-candidate", WorkspaceID: "wip-workspace", WorkflowID: "wip-workflow",
+		WorkflowStepID: "other-step", Title: "Candidate", State: v1.TaskStateCreated,
+	}
+	err := repo.UpdateTaskIfWorkflowStepHasCapacity(ctx, candidate, "wip-step", "wip-candidate", 1)
+	if err == nil || !errors.Is(err, wfmodels.ErrWIPLimitExceeded) {
+		t.Fatalf("error=%v, want typed WIP limit error", err)
+	}
+}
+
+func TestCreateTaskIfWorkflowStepHasCapacity_Concurrent(t *testing.T) {
+	repo, cleanup := createTestSQLiteRepo(t)
+	defer cleanup()
+
+	creator, ok := any(repo).(workflowStepCapacityCreator)
+	if !ok {
+		t.Fatal("task repository does not implement atomic workflow-step capacity creation")
+	}
+
+	const (
+		workerCount = 8
+		stepID      = "wip-step"
+	)
+	ctx := context.Background()
+	start := make(chan struct{})
+	results := make(chan error, workerCount)
+	var wg sync.WaitGroup
+	for i := 0; i < workerCount; i++ {
+		wg.Add(1)
+		go func(index int) {
+			defer wg.Done()
+			<-start
+			results <- creator.CreateTaskIfWorkflowStepHasCapacity(ctx, &models.Task{
+				ID:             fmt.Sprintf("wip-task-%d", index),
+				WorkspaceID:    "wip-workspace",
+				WorkflowID:     "wip-workflow",
+				WorkflowStepID: stepID,
+				Title:          fmt.Sprintf("Task %d", index),
+				State:          v1.TaskStateCreated,
+			}, stepID, 2)
+		}(i)
+	}
+	close(start)
+	wg.Wait()
+	close(results)
+
+	created, rejected := 0, 0
+	for err := range results {
+		if err == nil {
+			created++
+			continue
+		}
+		if !strings.Contains(strings.ToLower(err.Error()), "wip limit exceeded") {
+			t.Fatalf("unexpected create error: %v", err)
+		}
+		rejected++
+	}
+	if created != 2 || rejected != workerCount-2 {
+		t.Fatalf("created=%d rejected=%d, want created=2 rejected=%d", created, rejected, workerCount-2)
+	}
+
+	occupants, err := repo.CountTasksByWorkflowStep(ctx, stepID)
+	if err != nil {
+		t.Fatalf("count occupants: %v", err)
+	}
+	if occupants != 2 {
+		t.Fatalf("occupants=%d, want 2", occupants)
+	}
+}
diff --git a/apps/backend/internal/task/service/service.go b/apps/backend/internal/task/service/service.go
index 05cb2602a6..8df4fc8ced 100644
--- a/apps/backend/internal/task/service/service.go
+++ b/apps/backend/internal/task/service/service.go
@@ -145,6 +145,7 @@ type StartStepResolver interface {
 
 var (
 	ErrActiveTaskSessions        = errors.New("active agent sessions exist")
+	ErrWIPLimitExceeded          = wfmodels.ErrWIPLimitExceeded
 	ErrInvalidRepositorySettings = errors.New("invalid repository settings")
 	ErrInvalidExecutorConfig     = errors.New("invalid executor config")
 	// Workspace-source sentinels are the service boundary consumed by the HTTP
diff --git a/apps/backend/internal/task/service/service_tasks.go b/apps/backend/internal/task/service/service_tasks.go
index 20d94cd6a1..e3a4e3a77e 100644
--- a/apps/backend/internal/task/service/service_tasks.go
+++ b/apps/backend/internal/task/service/service_tasks.go
@@ -79,6 +79,10 @@ type taskEnvironmentOwnerTransferer interface {
 	TransferTaskEnvironmentToTask(ctx context.Context, envID, taskID string) error
 }
 
+type workflowStepCapacityTaskCreator interface {
+	CreateTaskIfWorkflowStepHasCapacity(ctx context.Context, task *models.Task, targetStepID string, limit int) error
+}
+
 // Task operations
 
 // isOfficeRequest returns true if the request should create an office task.
@@ -133,7 +137,7 @@ func (s *Service) CreateTask(ctx context.Context, req *CreateTaskRequest) (*mode
 		}
 	}
 
-	if err := s.tasks.CreateTask(ctx, task); err != nil {
+	if err := s.createTaskWithCapacity(ctx, task); err != nil {
 		s.logger.Error("failed to create task", zap.Error(err))
 		return nil, err
 	}
@@ -163,6 +167,30 @@ func (s *Service) CreateTask(ctx context.Context, req *CreateTaskRequest) (*mode
 	return task, nil
 }
 
+func (s *Service) createTaskWithCapacity(ctx context.Context, task *models.Task) error {
+	if task.IsEphemeral || task.WorkflowStepID == "" {
+		return s.tasks.CreateTask(ctx, task)
+	}
+	if s.workflowStepGetter == nil {
+		return s.tasks.CreateTask(ctx, task)
+	}
+	step, err := s.workflowStepGetter.GetStep(ctx, task.WorkflowStepID)
+	if err != nil {
+		return fmt.Errorf("load workflow step %s for task creation: %w", task.WorkflowStepID, err)
+	}
+	if step == nil {
+		return fmt.Errorf("%w: workflow step not found: %s", ErrInvalidTaskWorkflow, task.WorkflowStepID)
+	}
+	if step.WIPLimit <= 0 {
+		return s.tasks.CreateTask(ctx, task)
+	}
+	creator, ok := s.tasks.(workflowStepCapacityTaskCreator)
+	if !ok {
+		return fmt.Errorf("workflow step %s has WIP limit %d but capacity-aware task persistence is unavailable", step.ID, step.WIPLimit)
+	}
+	return creator.CreateTaskIfWorkflowStepHasCapacity(ctx, task, step.ID, step.WIPLimit)
+}
+
 // inheritParentRepositories fills req.Repositories from the parent task when a
 // subtask is created without explicit repositories. This applies to any
 // repo-less subtask (not only inherit_parent ones), matching the MCP
diff --git a/apps/backend/internal/task/service/service_tasks_wip_test.go b/apps/backend/internal/task/service/service_tasks_wip_test.go
new file mode 100644
index 0000000000..f39cbb2df3
--- /dev/null
+++ b/apps/backend/internal/task/service/service_tasks_wip_test.go
@@ -0,0 +1,121 @@
+package service
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"github.com/kandev/kandev/internal/task/models"
+	wfmodels "github.com/kandev/kandev/internal/workflow/models"
+	v1 "github.com/kandev/kandev/pkg/api/v1"
+)
+
+type fixedStartStepResolver struct {
+	stepID string
+}
+
+func (r fixedStartStepResolver) ResolveStartStep(context.Context, string) (string, error) {
+	return r.stepID, nil
+}
+
+func (r fixedStartStepResolver) ResolveFirstStep(context.Context, string) (string, error) {
+	return r.stepID, nil
+}
+
+func TestCreateTask_RejectsFullWIPStepBeforePersistence(t *testing.T) {
+	svc, events, repo := createTestService(t)
+	ctx := context.Background()
+	seedWIPWorkflow(t, ctx, repo)
+	svc.SetWorkflowStepGetter(&fakeWorkflowStepGetter{steps: map[string]*wfmodels.WorkflowStep{
+		"review-step": {ID: "review-step", WorkflowID: "wip-workflow", Name: "Review", WIPLimit: 1},
+	}})
+	if err := repo.CreateTask(ctx, &models.Task{
+		ID: "wip-occupant", WorkspaceID: "wip-workspace", WorkflowID: "wip-workflow",
+		WorkflowStepID: "review-step", Title: "Occupant", State: v1.TaskStateCreated,
+	}); err != nil {
+		t.Fatalf("seed occupant: %v", err)
+	}
+
+	_, err := svc.CreateTask(ctx, &CreateTaskRequest{
+		WorkspaceID: "wip-workspace", WorkflowID: "wip-workflow", WorkflowStepID: "review-step",
+		Title: "Rejected", Description: "must not persist",
+	})
+	if err == nil || !errors.Is(err, wfmodels.ErrWIPLimitExceeded) {
+		t.Fatalf("error=%v, want typed WIP limit rejection", err)
+	}
+	if _, err := repo.GetTask(ctx, "wip-occupant"); err != nil {
+		t.Fatalf("occupant disappeared: %v", err)
+	}
+	tasks, err := repo.ListTasksByWorkflowStep(ctx, "review-step")
+	if err != nil {
+		t.Fatalf("list step tasks: %v", err)
+	}
+	if len(tasks) != 1 {
+		t.Fatalf("step task count=%d, want 1", len(tasks))
+	}
+	if len(events.GetPublishedEvents()) != 0 {
+		t.Fatalf("published events=%d, want none", len(events.GetPublishedEvents()))
+	}
+}
+
+func TestCreateTask_ResolvedStartStepUsesWIPAdmission(t *testing.T) {
+	svc, _, repo := createTestService(t)
+	ctx := context.Background()
+	seedWIPWorkflow(t, ctx, repo)
+	svc.SetWorkflowStepGetter(&fakeWorkflowStepGetter{steps: map[string]*wfmodels.WorkflowStep{
+		"review-step": {ID: "review-step", WorkflowID: "wip-workflow", Name: "Review", WIPLimit: 1},
+	}})
+	svc.SetStartStepResolver(fixedStartStepResolver{stepID: "review-step"})
+	if err := repo.CreateTask(ctx, &models.Task{
+		ID: "wip-occupant-resolved", WorkspaceID: "wip-workspace", WorkflowID: "wip-workflow",
+		WorkflowStepID: "review-step", Title: "Occupant", State: v1.TaskStateCreated,
+	}); err != nil {
+		t.Fatalf("seed occupant: %v", err)
+	}
+
+	_, err := svc.CreateTask(ctx, &CreateTaskRequest{
+		WorkspaceID: "wip-workspace", WorkflowID: "wip-workflow",
+		Title: "Rejected resolved start", Description: "must not persist",
+	})
+	if err == nil || !errors.Is(err, wfmodels.ErrWIPLimitExceeded) {
+		t.Fatalf("error=%v, want typed WIP limit rejection", err)
+	}
+}
+
+func TestCreateTask_UnlimitedWIPStepPreservesCreation(t *testing.T) {
+	svc, _, repo := createTestService(t)
+	ctx := context.Background()
+	seedWIPWorkflow(t, ctx, repo)
+	svc.SetWorkflowStepGetter(&fakeWorkflowStepGetter{steps: map[string]*wfmodels.WorkflowStep{
+		"unlimited-step": {ID: "unlimited-step", WorkflowID: "wip-workflow", Name: "Unlimited"},
+	}})
+
+	for i := 0; i < 3; i++ {
+		if _, err := svc.CreateTask(ctx, &CreateTaskRequest{
+			WorkspaceID: "wip-workspace", WorkflowID: "wip-workflow", WorkflowStepID: "unlimited-step",
+			Title: "Unlimited task", Description: "allowed",
+		}); err != nil {
+			t.Fatalf("create task %d: %v", i, err)
+		}
+	}
+	occupants, err := repo.CountTasksByWorkflowStep(ctx, "unlimited-step")
+	if err != nil {
+		t.Fatalf("count occupants: %v", err)
+	}
+	if occupants != 3 {
+		t.Fatalf("occupants=%d, want 3", occupants)
+	}
+}
+
+func seedWIPWorkflow(t *testing.T, ctx context.Context, repo interface {
+	CreateWorkspace(context.Context, *models.Workspace) error
+	CreateWorkflow(context.Context, *models.Workflow) error
+}) {
+	t.Helper()
+	if err := repo.CreateWorkspace(ctx, &models.Workspace{ID: "wip-workspace", Name: "WIP Workspace"}); err != nil {
+		t.Fatalf("seed workspace: %v", err)
+	}
+	if err := repo.CreateWorkflow(ctx, &models.Workflow{ID: "wip-workflow", WorkspaceID: "wip-workspace", Name: "WIP Workflow"}); err != nil {
+		t.Fatalf("seed workflow: %v", err)
+	}
+}
diff --git a/apps/backend/internal/task/service/service_workflow.go b/apps/backend/internal/task/service/service_workflow.go
index cdf36e527b..d316ce8530 100644
--- a/apps/backend/internal/task/service/service_workflow.go
+++ b/apps/backend/internal/task/service/service_workflow.go
@@ -677,7 +677,7 @@ func (s *Service) validateMoveWIPLimit(ctx context.Context, task *models.Task, t
 		return fmt.Errorf("failed to count target workflow step tasks: %w", err)
 	}
 	if occupants >= targetStep.WIPLimit {
-		return fmt.Errorf("WIP limit exceeded for workflow step %s: limit %d already occupied", targetStep.ID, targetStep.WIPLimit)
+		return wfmodels.NewWIPLimitError(targetStep.ID, targetStep.WIPLimit, occupants)
 	}
 	return nil
 }
@@ -839,8 +839,7 @@ func (s *Service) validateBulkMoveWIPCapacity(ctx context.Context, tasks []*mode
 		return fmt.Errorf("failed to count target workflow step tasks: %w", err)
 	}
 	if occupants+incoming > targetStep.WIPLimit {
-		return fmt.Errorf("WIP limit exceeded for workflow step %s: limit %d, occupied %d, moving %d",
-			targetStep.ID, targetStep.WIPLimit, occupants, incoming)
+		return fmt.Errorf("%w: moving %d task(s)", wfmodels.NewWIPLimitError(targetStep.ID, targetStep.WIPLimit, occupants), incoming)
 	}
 	return nil
 }
diff --git a/apps/backend/internal/workflow/models/errors.go b/apps/backend/internal/workflow/models/errors.go
new file mode 100644
index 0000000000..0fc7830b80
--- /dev/null
+++ b/apps/backend/internal/workflow/models/errors.go
@@ -0,0 +1,30 @@
+package models
+
+import (
+	"errors"
+	"fmt"
+)
+
+// ErrWIPLimitExceeded identifies a workflow-step capacity rejection.
+var ErrWIPLimitExceeded = errors.New("workflow step WIP limit exceeded")
+
+// WIPLimitError carries workflow-step capacity details while preserving
+// errors.Is(err, ErrWIPLimitExceeded) classification at service boundaries.
+type WIPLimitError struct {
+	StepID   string
+	Limit    int
+	Occupied int
+}
+
+func (e *WIPLimitError) Error() string {
+	return fmt.Sprintf("WIP limit exceeded for workflow step %s: limit %d already occupied (%d)", e.StepID, e.Limit, e.Occupied)
+}
+
+func (e *WIPLimitError) Unwrap() error {
+	return ErrWIPLimitExceeded
+}
+
+// NewWIPLimitError constructs a typed capacity conflict for a workflow step.
+func NewWIPLimitError(stepID string, limit, occupied int) error {
+	return &WIPLimitError{StepID: stepID, Limit: limit, Occupied: occupied}
+}
diff --git a/apps/web/e2e/tests/kanban/mobile-kanban.spec.ts b/apps/web/e2e/tests/kanban/mobile-kanban.spec.ts
index eacea47822..2b516d17d7 100644
--- a/apps/web/e2e/tests/kanban/mobile-kanban.spec.ts
+++ b/apps/web/e2e/tests/kanban/mobile-kanban.spec.ts
@@ -574,7 +574,6 @@ test.describe("Mobile kanban view", () => {
       is_start_step: true,
     });
     await apiClient.createWorkflowStep(workflow.id, "Done", 1);
-    await apiClient.updateWorkflowStep(limitedStep.id, { wip_limit: 1 });
     await apiClient.createTask(seedData.workspaceId, "Mobile WIP One", {
       workflow_id: workflow.id,
       workflow_step_id: limitedStep.id,
@@ -583,6 +582,9 @@ test.describe("Mobile kanban view", () => {
       workflow_id: workflow.id,
       workflow_step_id: limitedStep.id,
     });
+    // Seed a legacy over-limit state by applying the limit after task creation;
+    // creation itself must now reject capacity overflow.
+    await apiClient.updateWorkflowStep(limitedStep.id, { wip_limit: 1 });
     await apiClient.saveUserSettings({
       workspace_id: seedData.workspaceId,
       workflow_filter_id: workflow.id,
diff --git a/docs/plans/wip-limited-task-creation/plan.md b/docs/plans/wip-limited-task-creation/plan.md
new file mode 100644
index 0000000000..047d40e3b6
--- /dev/null
+++ b/docs/plans/wip-limited-task-creation/plan.md
@@ -0,0 +1,199 @@
+---
+spec: docs/specs/tasks/wip-limit-pull-system.md
+created: 2026-07-27
+status: complete
+---
+
+# Implementation Plan: WIP-Limited Task Creation
+
+## Overview
+
+Task moves already enforce workflow-step WIP limits, but `CreateTask` inserts a
+task directly into its resolved step without capacity admission. GitHub review
+watches fan out one goroutine per newly observed PR and then auto-start every
+successfully created task, so a WIP-2 start step can receive and start all eight
+tasks from one poll.
+
+The fix adds a typed, transactionally enforced WIP admission primitive, uses it
+from task creation for both explicit and resolved start steps, maps capacity
+rejections consistently across HTTP, WebSocket, MCP, and watcher dispatch, and
+documents the corrected behavior. Watcher integration coverage also locks in
+the existing lifecycle contract: an admitted task remains in its configured
+step while the agent is starting and working, and advances only on a genuine
+turn-complete event.
+
+## Confirmed Root Cause
+
+- `github.Service.TriggerReviewWatch` publishes an event for every new PR.
+- `orchestrator.Service.handleNewReviewPR` starts one background
+  `createReviewTask` goroutine per event.
+- `task/service.Service.CreateTask` calls `TaskRepository.CreateTask` without
+  reading the resolved step's `WIPLimit`.
+- WIP validation exists only on move/transition paths.
+- `createReviewTask` calls `StartTask` for every task created in an
+  `auto_start_agent` step.
+
+The smallest reproduction is an empty auto-start workflow start step with
+`wip_limit: 2` receiving eight concurrent non-ephemeral task creations. Current
+behavior persists eight tasks; the required behavior admits exactly two.
+
+---
+
+## Backend
+
+### Typed WIP capacity contract
+
+- Add `ErrWIPLimitExceeded` to
+  `apps/backend/internal/workflow/models/errors.go` so task repository, task
+  service, HTTP/WS/MCP adapters, and orchestrator code classify capacity
+  without parsing error strings.
+- Preserve user-facing context by wrapping the sentinel with the step ID,
+  configured limit, and current occupancy.
+
+### Atomic repository admission
+
+- Extend the narrow task-placement repository contract with
+  `CreateTaskIfWorkflowStepHasCapacity(ctx, task, targetStepID, limit)`.
+- Implement it in
+  `apps/backend/internal/task/repository/sqlite/task.go` using the same task
+  insert and runner-participant transaction as `CreateTask`.
+- Serialize admission on the target `workflow_steps` row. PostgreSQL uses
+  `SELECT ... FOR UPDATE`; SQLite already uses a single writer connection, so
+  the count and insert remain in one serialized write transaction.
+- Count only non-archived, non-ephemeral occupants, matching existing WIP move
+  semantics.
+- Reuse the step-capacity lock and typed sentinel from
+  `UpdateTaskIfWorkflowStepHasCapacity` so creation, moves, and pulls share one
+  concurrency rule.
+
+### Task service enforcement
+
+- In `task/service.Service.CreateTask`, resolve and validate the final workflow
+  step before persistence.
+- For a positive `WIPLimit`, call the capacity-aware repository method instead
+  of the unconditional insert. This includes requests that omit
+  `workflow_step_id` and resolve to the workflow start step.
+- Fail closed when a positive WIP limit is configured but the repository does
+  not implement capacity-aware creation.
+- Leave `wip_limit: 0`, workflow-less ephemeral tasks, and existing unlimited
+  task creation unchanged.
+- Ensure a capacity rejection occurs before blockers, repositories,
+  `task.created`, task sessions, or auto-start side effects.
+
+### Caller error behavior
+
+- HTTP task creation returns `409 Conflict` with the WIP error.
+- WebSocket and MCP task creation return `ErrorCodeConflict` with the same
+  message.
+- GitHub review dispatch recognizes `ErrWIPLimitExceeded` as a deferral:
+  release the review-PR reservation, do not attach a task ID or auto-start, and
+  log at a non-error level so later polls can retry without false failure
+  noise.
+- Other task-creation errors retain their current handling.
+
+---
+
+## Tests
+
+- **What:** concurrent repository admission never exceeds WIP capacity.
+  **File:** `apps/backend/internal/task/repository/task_wip_test.go`.
+  **How:** seed a WIP-2 step, synchronize eight goroutines, assert exactly two
+  inserts and six `errors.Is(err, ErrWIPLimitExceeded)` results; run with
+  `-race`.
+- **What:** explicit and resolved start-step creation obey WIP, while
+  `wip_limit: 0` remains unlimited and rejected creation emits no
+  `task.created`.
+  **File:** new
+  `apps/backend/internal/task/service/service_tasks_wip_test.go`.
+  **How:** SQLite-backed service integration tests using real workflow/task
+  repositories.
+- **What:** HTTP, WebSocket, and MCP task creation classify WIP rejection as a
+  conflict.
+  **Files:** task handler and MCP handler tests.
+  **How:** focused adapter tests asserting status/error code and message.
+- **What:** review-watch capacity rejection releases the PR reservation and
+  never assigns a task ID or auto-starts.
+  **File:**
+  `apps/backend/internal/orchestrator/event_handlers_github_review_test.go`.
+  **How:** deterministic fake creator returning the typed WIP error, followed
+  by a retry that succeeds.
+- **What:** an admitted watcher task does not advance its workflow during
+  startup or active work.
+  **File:**
+  `apps/backend/internal/orchestrator/event_handlers_github_review_test.go`.
+  **How:** create directly in an auto-start `Review` step, assert the task
+  remains there after creation and boot-ready, then emit a real turn-complete
+  event and assert one transition to `Done`.
+
+## Frontend
+
+No frontend code changes are required. Existing task-creation surfaces already
+display backend errors; the backend adapters must return a conflict instead of
+an internal error.
+
+## E2E Tests
+
+No browser E2E is planned because there is no visual or interaction change.
+The concurrency invariant and every external error boundary are covered by
+targeted Go tests.
+
+## Public Documentation
+
+- Update `docs/public/tasks-and-workflows.md` and
+  `docs/public/workflow-tips.md` to state that WIP applies to initial task
+  creation, including integration-created tasks, and that full-step creation
+  is rejected for later retry.
+- Validate the public docs after the wording changes.
+
+## Implementation Waves And Parallel Candidates
+
+All tasks are sequential because each consumes the typed contract or behavior
+introduced by the preceding task.
+
+- [x] [Task 01: Atomic repository admission](task-01-atomic-repository-admission.md)
+- [x] [Task 02: Task-service WIP enforcement](task-02-task-service-enforcement.md)
+- [x] [Task 03: Conflict adapters and review-watch deferral](task-03-conflict-adapters-and-watcher-deferral.md)
+- [x] [Task 04: Public documentation](task-04-public-documentation.md)
+
+## Completion Evidence
+
+- Repository admission is transactionally serialized, with PostgreSQL step-row
+  locking and SQLite writer serialization; focused repository tests pass.
+- Task creation enforces explicit and resolved workflow-step WIP limits, while
+  unlimited and ephemeral creation remain compatible; focused service tests
+  pass.
+- HTTP returns `409 Conflict`; WebSocket and MCP return conflict error codes;
+  watcher capacity failures release reservations and skip assignment/start.
+- Review watcher lifecycle coverage confirms boot-ready does not advance the
+  task and the first genuine turn completion moves Review to Done once.
+- Public documentation validation: `node scripts/validate-public-docs.test.mjs`
+  (58/58 passed).
+- Backend verification: `go test -tags fts5 ./internal/task/repository
+  ./internal/task/service ./internal/task/handlers ./internal/mcp/handlers
+  ./internal/orchestrator -count=1` (2,562 tests passed).
+
+No task is marked `parallel-safe`; waves do not authorize subagent execution.
+
+## Risks
+
+- A pre-insert count outside the write transaction would preserve the original
+  race. Admission must be serialized per target step in both SQLite and
+  PostgreSQL.
+- Rejected review PRs must release their dedup reservations or they will never
+  retry.
+- Watcher auto-start coverage must distinguish `agent.boot_ready` from a real
+  turn-complete event; treating startup readiness as completion would move a
+  newly admitted task out of the limited step before its review runs.
+- The resolved start-step path must be covered explicitly; checking only
+  request-supplied `workflow_step_id` would leave the reported configuration
+  broken.
+- Existing tests intentionally seed over-limit UI states through task creation.
+  Those fixtures must use repository-level setup or another explicit legacy
+  state mechanism rather than weakening the production invariant.
+
+## Out of Scope
+
+- A GitHub review-watch `max_inflight_tasks` setting.
+- Profile-wide session concurrency limits.
+- Automatic feeder-step creation or watcher configuration changes.
+- UI redesign or new WIP controls.
diff --git a/docs/plans/wip-limited-task-creation/task-01-atomic-repository-admission.md b/docs/plans/wip-limited-task-creation/task-01-atomic-repository-admission.md
new file mode 100644
index 0000000000..3693412fe0
--- /dev/null
+++ b/docs/plans/wip-limited-task-creation/task-01-atomic-repository-admission.md
@@ -0,0 +1,65 @@
+---
+id: "01-atomic-repository-admission"
+title: "Atomic repository admission"
+status: done
+wave: 1
+depends_on: []
+plan: "plan.md"
+spec: "../../specs/tasks/wip-limit-pull-system.md"
+---
+
+# Task 01: Atomic Repository Admission
+
+## Acceptance
+
+- The task repository can create a task only when the target workflow step has
+  capacity, counting active non-ephemeral occupants with the same semantics as
+  WIP-limited moves.
+- Eight synchronized creates against an empty WIP-2 step persist exactly two
+  tasks; every rejected call matches the typed WIP sentinel.
+- PostgreSQL admission locks the target workflow-step row for the transaction;
+  SQLite admission remains serialized through its single writer.
+- Existing capacity-aware updates return the same typed WIP error.
+
+## Verification
+
+```bash
+cd apps/backend
+go test -tags fts5 -race ./internal/task/repository/... -run 'Test.*WorkflowStep.*Capacity|Test.*WIP.*Concurrent' -count=1
+```
+
+## Files likely touched
+
+- `apps/backend/internal/workflow/models/errors.go`
+- `apps/backend/internal/task/repository/interface.go`
+- `apps/backend/internal/task/repository/sqlite/task.go`
+- `apps/backend/internal/task/repository/task_repository_test.go`
+- `apps/backend/internal/task/repository/sqlite/task_order_test.go` or a focused
+  dialect test if the PostgreSQL lock query is extracted
+
+## Dependencies
+
+None.
+
+## Parallelism
+
+`sequential`
+
+## Inputs
+
+- Spec sections `What`, `Failure Modes`, and `Scenarios`.
+- Existing `UpdateTaskIfWorkflowStepHasCapacity` transaction.
+- SQLite writer configuration in `apps/backend/internal/db/sqlite.go`.
+- PostgreSQL dialect detection in `apps/backend/internal/db/dialect`.
+
+## Output contract
+
+Mark this task `in_progress` before the RED test and `done` after GREEN and
+refactor. Update `plan.md` in the same conversation and report the failing-test
+evidence, repository API/transaction changes, files changed, exact test result,
+and remaining database portability risks.
+
+## Evidence
+
+`go test ./internal/task/repository -run 'Test(CreateTaskIfWorkflowStepHasCapacity|UpdateTaskIfWorkflowStepHasCapacity)' -count=1`
+passed (2 tests).
diff --git a/docs/plans/wip-limited-task-creation/task-02-task-service-enforcement.md b/docs/plans/wip-limited-task-creation/task-02-task-service-enforcement.md
new file mode 100644
index 0000000000..8652830393
--- /dev/null
+++ b/docs/plans/wip-limited-task-creation/task-02-task-service-enforcement.md
@@ -0,0 +1,63 @@
+---
+id: "02-task-service-enforcement"
+title: "Task-service WIP enforcement"
+status: done
+wave: 2
+depends_on: ["01-atomic-repository-admission"]
+plan: "plan.md"
+spec: "../../specs/tasks/wip-limit-pull-system.md"
+---
+
+# Task 02: Task-Service WIP Enforcement
+
+## Acceptance
+
+- `Service.CreateTask` uses atomic capacity admission for a positive WIP limit
+  after resolving the final workflow step.
+- Explicit and automatically resolved start-step creation reject a full step
+  with the typed WIP error and persist no task, repository, blocker, event, or
+  session side effect.
+- Unlimited workflow steps and workflow-less ephemeral tasks preserve current
+  behavior.
+- A positive WIP limit fails closed if capacity-aware persistence is
+  unavailable.
+
+## Verification
+
+```bash
+cd apps/backend
+go test -tags fts5 ./internal/task/service -run 'TestCreateTask.*WIP|TestCreateTask.*StartStep|TestCreateTask.*Unlimited' -count=1
+```
+
+## Files likely touched
+
+- `apps/backend/internal/task/service/service_tasks.go`
+- `apps/backend/internal/task/service/service_tasks_wip_test.go`
+- `apps/backend/internal/task/service/service.go` only if a narrow shared
+  interface or helper belongs there
+
+## Dependencies
+
+Task 01.
+
+## Parallelism
+
+`sequential`
+
+## Inputs
+
+- Task 01's typed error and capacity-aware repository method.
+- `resolveWorkflowStep`, `validateTaskWorkflow`, and `buildTask` in
+  `service_tasks.go`.
+- Existing task-event test helpers in `service_events_test.go`.
+
+## Output contract
+
+Mark this task `in_progress` before the RED test and `done` after GREEN and
+refactor. Update `plan.md` and report explicit/resolved-step coverage,
+side-effect assertions, files changed, exact test result, blockers, and risks.
+
+## Evidence
+
+Explicit-step, resolved-start-step, unlimited, and no-event rejection tests
+pass in `internal/task/service`.
diff --git a/docs/plans/wip-limited-task-creation/task-03-conflict-adapters-and-watcher-deferral.md b/docs/plans/wip-limited-task-creation/task-03-conflict-adapters-and-watcher-deferral.md
new file mode 100644
index 0000000000..e443e8a144
--- /dev/null
+++ b/docs/plans/wip-limited-task-creation/task-03-conflict-adapters-and-watcher-deferral.md
@@ -0,0 +1,78 @@
+---
+id: "03-conflict-adapters-and-watcher-deferral"
+title: "Conflict adapters and review-watch deferral"
+status: done
+wave: 3
+depends_on: ["02-task-service-enforcement"]
+plan: "plan.md"
+spec: "../../specs/tasks/wip-limit-pull-system.md"
+---
+
+# Task 03: Conflict Adapters and Review-Watch Deferral
+
+## Acceptance
+
+- HTTP task creation returns `409 Conflict` for the typed WIP error.
+- WebSocket and MCP task creation return `ErrorCodeConflict` and preserve the
+  actionable step/limit message.
+- GitHub review task creation treats WIP rejection as deferred capacity:
+  releases the PR reservation, assigns no task ID, starts no task, and permits a
+  later retry to succeed.
+- An admitted GitHub review task created directly in an auto-start `Review`
+  step remains in that step during startup and active work; `agent.boot_ready`
+  does not advance it, and the first genuine turn completion moves it to the
+  configured next step exactly once.
+- Non-capacity task-creation failures retain their current error behavior.
+
+## Verification
+
+```bash
+cd apps/backend
+go test -tags fts5 ./internal/task/handlers -run 'Test.*CreateTask.*WIP' -count=1
+go test -tags fts5 ./internal/mcp/handlers -run 'TestHandleCreateTask.*WIP' -count=1
+go test -tags fts5 ./internal/orchestrator -run 'Test(CreateReviewTask_.*|GitHubWatcherTaskCreateError|ReviewWatchLifecycle_.*)' -count=1
+```
+
+## Files likely touched
+
+- `apps/backend/internal/task/handlers/errors.go`
+- `apps/backend/internal/task/handlers/task_http_handlers.go`
+- `apps/backend/internal/task/handlers/task_http_handlers_test.go`
+- `apps/backend/internal/task/handlers/task_ws_handlers.go`
+- `apps/backend/internal/task/handlers/task_ws_handlers_test.go`
+- `apps/backend/internal/mcp/handlers/handlers.go`
+- `apps/backend/internal/mcp/handlers/handlers_test.go`
+- `apps/backend/internal/orchestrator/event_handlers_github.go`
+- `apps/backend/internal/orchestrator/event_handlers_github_review_test.go`
+
+## Dependencies
+
+Task 02.
+
+## Evidence
+
+HTTP, WebSocket, and MCP focused tests pass. GitHub review capacity rejection
+coverage confirms reservation release without task assignment or auto-start;
+the lifecycle regression confirms boot-ready stays in Review and the first
+genuine turn completion advances once.
+
+## Parallelism
+
+`sequential`
+
+## Inputs
+
+- Task 02's service error behavior.
+- Existing `isMoveConflict`, `isTaskCreateValidationError`, and WebSocket error
+  mappings.
+- Existing review reservation release test in
+  `event_handlers_github_review_test.go`.
+- Existing boot-ready/turn-complete separation in
+  `event_handlers_agent.go` and its regression tests.
+
+## Output contract
+
+Mark this task `in_progress` before the RED tests and `done` after GREEN and
+refactor. Update `plan.md` and report each adapter's observed error contract,
+review reservation/retry evidence, watcher lifecycle-ordering evidence, files
+changed, exact test results, blockers, and risks.
diff --git a/docs/plans/wip-limited-task-creation/task-04-public-documentation.md b/docs/plans/wip-limited-task-creation/task-04-public-documentation.md
new file mode 100644
index 0000000000..b4eebb181c
--- /dev/null
+++ b/docs/plans/wip-limited-task-creation/task-04-public-documentation.md
@@ -0,0 +1,56 @@
+---
+id: "04-public-documentation"
+title: "Public WIP documentation"
+status: done
+wave: 4
+depends_on: ["03-conflict-adapters-and-watcher-deferral"]
+plan: "plan.md"
+spec: "../../specs/tasks/wip-limit-pull-system.md"
+---
+
+# Task 04: Public WIP Documentation
+
+## Acceptance
+
+- Public workflow documentation states that WIP capacity applies to initial
+  task creation as well as moves and transitions.
+- Documentation explains that integration-created tasks rejected for capacity
+  remain deferred for later retry and do not auto-start.
+- Public-doc validation succeeds.
+
+## Evidence
+
+Updated `docs/public/tasks-and-workflows.md` and
+`docs/public/workflow-tips.md`. `node scripts/validate-public-docs.test.mjs`
+passed all 58 tests.
+
+## Verification
+
+```bash
+node --test scripts/validate-public-docs.test.mjs
+node scripts/validate-public-docs.mjs
+```
+
+## Files likely touched
+
+- `docs/public/tasks-and-workflows.md`
+- `docs/public/workflow-tips.md`
+
+## Dependencies
+
+Task 03.
+
+## Parallelism
+
+`sequential`
+
+## Inputs
+
+- Amended WIP spec.
+- Final typed conflict and review-watch deferral behavior from Tasks 01-03.
+
+## Output contract
+
+Mark this task `in_progress` before editing and `done` after validation. Update
+`plan.md` and report the public wording changed, files changed, exact validation
+results, blockers, and residual documentation risks.
diff --git a/docs/public/tasks-and-workflows.md b/docs/public/tasks-and-workflows.md
index ef2318bea1..224cfb6c25 100644
--- a/docs/public/tasks-and-workflows.md
+++ b/docs/public/tasks-and-workflows.md
@@ -182,6 +182,19 @@ New steps allow manual moves by default. **Show in command panel** also defaults
 | **WIP limit**             | Maximum active, non-archived, non-ephemeral tasks accepted by the step. `0` means unlimited. A move into a full step is rejected; reordering within the same step does not consume another slot. |
 | **Pull from**             | Optional feeder step. When capacity opens, Kandev pulls eligible work from that step. It requires a nonzero WIP limit.                                                                           |
 
+The WIP check also applies when a task is created. It runs for an explicit
+`workflow_step_id` and for the workflow's resolved start step, and the
+admission check is atomic so concurrent creates cannot collectively exceed the
+limit. A full step returns a conflict that includes the step and limit; the
+rejected task, session, and task-created event are not persisted. Ephemeral
+tasks are not counted.
+
+Integration watchers use the same admission rule. For example, a GitHub review
+watch targeting a `Review` step with a limit of two admits at most two newly
+observed pull requests at a time. Pull requests that lose the capacity race
+remain eligible for a later poll; Kandev releases their temporary watch
+reservation and does not start an agent for them.
+
 Auto-archive is checked on a five-minute background interval and uses the task's last update time. Any task update postpones eligibility, so the archive is not guaranteed at the exact configured minute. Auto-archive affects the task itself, not its children.
 
 Pull configuration rejects self-references, cycles, and cross-workflow feeders. Pulling runs when a task vacates the limited step and fills each newly open slot. Candidates are ordered by board position, then priority (`critical`, `high`, `medium`, `low`, `none`), creation time, and ID. A candidate whose move fails—for example because its session is running or starting—is skipped for that pull pass.
@@ -200,6 +213,12 @@ Generic comment, blocker-resolution, approval, heartbeat, budget, and error trig
 
 When **On Turn Complete** moves a task, **Wait for agent completion signal** is available. With it enabled, a bare turn end leaves the task waiting; the agent must call `step_complete_kandev`. The call requires a summary and can include a handoff or blockers. It is idempotent within the step, runs asynchronously, and a user message sent before the transition is applied cancels that pending signal. Without the option, turn end counts as completion.
 
+An auto-started task stays in its current step while the agent session boots and
+while its first turn is running. A boot-ready event is not a turn completion.
+For example, a review step with `on_enter: auto_start_agent` and
+`on_turn_complete: move_to_next` moves to the next step only after the genuine
+review turn completes, not during startup.
+
 Plan mode can be disabled when the turn completes and/or when the task exits the step. A step prompt is Markdown and can include `{{task_prompt}}` to insert the original task description.
 
 ### Build a human gate
diff --git a/docs/public/workflow-tips.md b/docs/public/workflow-tips.md
index f3afdd7590..b4fd2bdfd9 100644
--- a/docs/public/workflow-tips.md
+++ b/docs/public/workflow-tips.md
@@ -70,6 +70,21 @@ Choose this for a deliberate multi-pass delivery process. It is excessive for sm
 
 Choose this for a local first-pass review. For repository-provider watch automation, use the relevant integration instead.
 
+### GitHub review watches and WIP limits
+
+A GitHub Review Watch creates each matching pull request in its configured
+workflow step. If that step has a WIP limit, admission is atomic: only tasks
+that obtain a slot are created. An admitted task auto-starts only when its
+configured workflow step enables `auto_start_agent`. A pull request rejected
+because the step is full is deferred for a later poll; its temporary watch
+reservation is released, so it is not lost or permanently marked as handled.
+
+If the target step has `auto_start_agent` on entry and `move_to_next` on turn
+completion, the task remains in the target step during agent startup and active
+review. `agent.boot_ready` only makes the session ready; the first real turn
+completion performs the configured move once. Use a human-gate transition when
+reviews must wait for approval instead of advancing automatically.
+
 ## Build a custom workflow
 
 Choose **Add Workflow**, give it a name, select **Custom**, and save it. Expand each step to edit its behavior. Reorder steps by dragging them; transition actions that say “next” or “previous” follow the saved position order.
diff --git a/docs/specs/tasks/wip-limit-pull-system.md b/docs/specs/tasks/wip-limit-pull-system.md
index 32a75219d1..389e42989a 100644
--- a/docs/specs/tasks/wip-limit-pull-system.md
+++ b/docs/specs/tasks/wip-limit-pull-system.md
@@ -1,5 +1,5 @@
 ---
-status: building
+status: shipped
 ---
 
 # WIP Limit Pull System
@@ -11,9 +11,23 @@ Workflow steps can define a work-in-progress limit and an optional feeder step.
 and preserves existing behavior. `pull_from_step_id` is empty by default; when it
 is set, it must reference another step in the same workflow.
 
-When a limited step is full, manual moves, drag/drop moves, API moves, MCP moves,
-bulk moves, and workflow-engine transitions into that step are rejected instead
-of overfilling the step. Same-step reordering is allowed.
+When a limited step is full, new task creation, manual moves, drag/drop moves,
+API moves, MCP moves, bulk moves, and workflow-engine transitions into that
+step are rejected instead of overfilling the step. The creation rule applies
+whether the caller names the step explicitly or Kandev resolves it as the
+workflow's start step. Same-step reordering is allowed.
+
+Capacity admission is atomic per workflow step. Concurrent task creation and
+move attempts cannot collectively admit more active, non-archived,
+non-ephemeral tasks than the configured limit.
+
+Integration watchers use the same admission rule as interactive and API task
+creation. If a GitHub review watch observes more pull requests than its target
+step can accept, Kandev creates only the tasks that obtain capacity and
+auto-starts those tasks only when the configured step enables
+`auto_start_agent`. Pull requests rejected for capacity remain eligible for a
+later poll; Kandev does not leave a task, session, repository association, or
+permanent review-watch reservation for a rejected attempt.
 
 When a task leaves a limited step that has a feeder configured, Kandev attempts
 to pull queued tasks from the feeder into the vacated step until the step reaches
@@ -52,13 +66,44 @@ ID.
 Moving into a full limited step returns a conflict with a user-visible message
 that includes the target step and limit. Optimistic UI moves must roll back.
 
+Creating a task in a full limited step returns the same conflict classification
+and does not persist any part of the rejected task. Internal automation treats
+the conflict as deferred work rather than an integration failure.
+
 If a pull attempt races with another actor that fills the slot, the pull attempt
 stops without overfilling the target step.
 
 Deleting a step clears any `pull_from_step_id` that points at it.
 
-## Scope Note
+## Scenarios
 
-`agent_profiles.max_concurrent_sessions` is related but separate. This feature
-adds per-step WIP limits and pull behavior. Kanban enforcement of profile-level
-session caps remains tracked separately unless implemented in the same change.
+- **GIVEN** a workflow step with `wip_limit: 2` and two active tasks, **WHEN** a
+  user, API client, MCP client, or integration watcher creates another
+  non-ephemeral task directly in that step, **THEN** creation is rejected as a
+  WIP-capacity conflict and no task or session is persisted.
+- **GIVEN** a workflow whose limited start step has `wip_limit: 2`, **WHEN** a
+  caller creates a task without specifying `workflow_step_id` and the start
+  step is full, **THEN** creation is rejected by the same capacity rule.
+- **GIVEN** an empty auto-start step with `wip_limit: 2`, **WHEN** a GitHub
+  review watch concurrently dispatches eight newly observed pull requests,
+  **THEN** exactly two review tasks obtain capacity and auto-start, while the
+  other six pull requests retain no task or reservation and remain eligible for
+  a later poll.
+- **GIVEN** a review watch targets a `Review` step with
+  `on_enter: auto_start_agent` and `on_turn_complete: move_to_next`, **WHEN** an
+  admitted review task starts, **THEN** it remains in `Review` throughout agent
+  startup and the active turn, ignores boot-ready events for workflow
+  advancement, and moves to the next step exactly once only after the real
+  agent turn completes.
+- **GIVEN** the review-watch scenario above and one accepted review task later
+  leaves the limited step, **WHEN** the watch polls again, **THEN** one
+  previously deferred pull request can obtain the newly available capacity.
+- **GIVEN** a step with `wip_limit: 0`, **WHEN** callers create tasks
+  concurrently in that step, **THEN** creation remains unlimited.
+
+## Out of scope
+
+- WIP limits do not replace `agent_profiles.max_concurrent_sessions` or impose
+  a profile-wide execution budget.
+- This change does not add a separate GitHub review-watch
+  `max_inflight_tasks` setting.