feat(orchestrator): surface stalled Office tasks without reclaiming them (#3279)

Co-authored-by: nova28 <17953305+nova28@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Carlos Florencio <carlosmflorencio@gmail.com>

AGENTS.md

@@ -174,6 +174,7 @@ history and remains immutable.
 
 - In dev mode (`KANDEV_MOCK_AGENT=true` or `debug.pprofEnabled`), `/debug/vars` exposes the stdlib expvar handler. Office provider-routing metrics live under `routing_*` (route attempts, fallbacks, parked runs, provider degraded/recovered counters). The metrics are also still emitted as structured `routing.metric.*` zap logs for human debugging.
 - ADR 0015's step-completion-signal telemetry lives under `workflow_*`: `workflow_step_completion_signal_received_total` (`internal/workflow/signalmetrics/`), labelled by `source` and `agent_type`, counts accepted `step_complete_kandev` signals. Its separate `workflow_step_completion_signal_fallback_used_total` counter is labelled by `agent_type` and counts manual fallback uses. The fallback button does not have a production increment site yet, so the counter remains zero until that UI ships.
+- Office stall detection (REQ-OFFICE-STALL-VISIBILITY) lives under `office_stall_*`: `office_stall_stranded_signal_total` (labelled by `gate`, which names which of the two watchdog gate sites saw it), `office_stall_decision_waiting_total`, and `office_stall_detector_skipped_total` (labelled by `reason`, so a detector that fails closed is visible rather than silent). Detection only: these counters never accompany a transition, a synthesized decision, or a queued run, because Office tasks are surfaced and never reclaimed. Also emitted as structured zap logs.
 
 ### GitHub Operations
 

apps/backend/internal/backendapp/main.go

@@ -1623,6 +1623,10 @@ func wireWorkflowEngineForOffice(
 	// REQ-OFFICE-REVIEW-SEATS-004.3: drop a required seat whose agent profile
 	// was deleted after casting, rather than waiting forever on it.
 	orchestratorSvc.SetEngineAgentProfileResolver(officeengineadapters.NewAgentProfileResolverAdapter(repos.Office))
+	// REQ-OFFICE-STALL-VISIBILITY-002: the decision-waiting detector's
+	// false-positive guard. Without this the detector reports nothing rather
+	// than reporting tasks it cannot prove are idle.
+	orchestratorSvc.SetOfficeRunInFlightReader(repos.Office)
 	eng := orchestratorSvc.WorkflowEngine()
 	if eng == nil {
 		log.Warn("workflow engine not initialised; office engine dispatcher disabled")

apps/backend/internal/office/repository/sqlite/runs_inflight.go

@@ -0,0 +1,52 @@
+package sqlite
+
+import (
+	"context"
+	"fmt"
+
+	"github.com/kandev/kandev/internal/db/dialect"
+	"github.com/kandev/kandev/internal/office/models"
+)
+
+// inFlightRunStatuses are the run statuses that mean work is still owed on a
+// task. The runs table has no "running" status: a run is `queued` until the
+// scheduler claims it and `claimed` while it executes, reaching `finished`,
+// `failed` or `cancelled` only at the end. Anything outside this pair is
+// terminal.
+var inFlightRunStatuses = []string{
+	string(models.RunStatusQueued),
+	string(models.RunStatusClaimed),
+}
+
+// HasInFlightRunForTask reports whether the task has a queued or claimed run.
+//
+// The runs table carries no task_id column — a run's task lives in its payload
+// JSON — so the predicate reads it out with dialect.JSONExtract rather than a
+// literal json_extract. Postgres is a supported driver
+// (internal/persistence/provider.go), and a SQLite-only expression here would
+// pass every SQLite test and fail at runtime on Postgres. Precedent:
+// HasPriorTasklessFailedRun in failure.go, covered by failure_postgres_test.go.
+//
+// Used by the Office decision-waiting detector as its false-positive guard: a
+// task with work in flight is being worked on, not stalled, however long it has
+// carried an undecided review seat.
+func (r *Repository) HasInFlightRunForTask(ctx context.Context, taskID string) (bool, error) {
+	if taskID == "" {
+		return false, nil
+	}
+	taskIDExpr := dialect.JSONExtract(r.ro.DriverName(), "payload", "task_id")
+	query := fmt.Sprintf(`
+		SELECT EXISTS (
+			SELECT 1 FROM runs
+			WHERE %s = ?
+			  AND status IN (?, ?)
+		)
+	`, taskIDExpr)
+	var exists bool
+	if err := r.ro.QueryRowxContext(
+		ctx, r.ro.Rebind(query), taskID, inFlightRunStatuses[0], inFlightRunStatuses[1],
+	).Scan(&exists); err != nil {
+		return false, fmt.Errorf("check in-flight run for task: %w", err)
+	}
+	return exists, nil
+}

apps/backend/internal/office/repository/sqlite/runs_inflight_postgres_test.go

@@ -0,0 +1,69 @@
+package sqlite_test
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/kandev/kandev/internal/office/repository/sqlite"
+	taskrepo "github.com/kandev/kandev/internal/task/repository/sqlite"
+	"github.com/kandev/kandev/internal/testutil"
+)
+
+// TestPostgresHasInFlightRunForTask is the PostgreSQL twin of
+// TestHasInFlightRunForTask. HasInFlightRunForTask reads task_id out of the
+// runs payload, and a literal json_extract(...) there is a syntax error on
+// Postgres (payload->>'task_id' is the Postgres form) — the same trap
+// TestPostgresHasPriorTasklessFailedRun was written for. Running the predicate
+// against a real Postgres backend makes a dialect regression fail loudly
+// instead of only in production, where its failure mode is a detector that
+// errors on every tick and surfaces nothing.
+// Skips unless KANDEV_TEST_POSTGRES_DSN is set.
+func TestPostgresHasInFlightRunForTask(t *testing.T) {
+	db := testutil.OpenIsolatedPostgres(t, testutil.PostgresDSNFromEnv(t))
+	ctx := context.Background()
+
+	// runs is created by the task repository's schema init, mirroring
+	// production boot order (see failure_postgres_test.go).
+	if _, err := taskrepo.NewWithDB(db, db, nil); err != nil {
+		t.Fatalf("init task repo: %v", err)
+	}
+	repo, err := sqlite.NewWithDB(db, db, nil)
+	if err != nil {
+		t.Fatalf("init office repo: %v", err)
+	}
+
+	now := time.Now().UTC()
+	cases := []struct {
+		status string
+		want   bool
+	}{
+		{"queued", true},
+		{"claimed", true},
+		{"finished", false},
+		{"failed", false},
+		{"cancelled", false},
+	}
+	for _, tc := range cases {
+		taskID := "pg-inflight-" + tc.status
+		seedPostgresRun(t, ctx, repo, "pg-run-"+tc.status, "agent-a", tc.status, now,
+			`{"task_id":"`+taskID+`"}`)
+
+		got, err := repo.HasInFlightRunForTask(ctx, taskID)
+		if err != nil {
+			t.Fatalf("HasInFlightRunForTask(%s): %v", tc.status, err)
+		}
+		if got != tc.want {
+			t.Errorf("HasInFlightRunForTask(status=%s) = %v, want %v", tc.status, got, tc.want)
+		}
+	}
+
+	// Scoping: the in-flight rows above must not answer for a different task.
+	got, err := repo.HasInFlightRunForTask(ctx, "pg-inflight-absent")
+	if err != nil {
+		t.Fatalf("HasInFlightRunForTask (absent task): %v", err)
+	}
+	if got {
+		t.Error("got true, want false — no run carries this task_id")
+	}
+}

apps/backend/internal/office/repository/sqlite/runs_inflight_test.go

@@ -0,0 +1,102 @@
+package sqlite_test
+
+import (
+	"context"
+	"testing"
+	"time"
+)
+
+// HasInFlightRunForTask is the false-positive guard of the Office
+// decision-waiting detector (REQ-OFFICE-STALL-VISIBILITY-002): a task with a
+// queued or claimed run is being worked on, so it must never be surfaced as
+// waiting on a decision. These tests pin both halves of that predicate — the
+// two in-flight statuses answer true, every terminal status answers false —
+// because a guard that silently answers false for `claimed` turns the whole
+// detector into a false-positive generator.
+
+func TestHasInFlightRunForTask(t *testing.T) {
+	repo := newTestRepo(t)
+	ctx := context.Background()
+
+	cases := []struct {
+		status string
+		want   bool
+	}{
+		{"queued", true},
+		{"claimed", true},
+		{"finished", false},
+		{"failed", false},
+		{"cancelled", false},
+	}
+	for _, tc := range cases {
+		t.Run(tc.status, func(t *testing.T) {
+			taskID := "task-" + tc.status
+			finishedAt := time.Now().UTC()
+			var stamp *time.Time
+			if !tc.want {
+				stamp = &finishedAt
+			}
+			seedCancelRun(t, repo, taskID, tc.status, nil, stamp)
+
+			got, err := repo.HasInFlightRunForTask(ctx, taskID)
+			if err != nil {
+				t.Fatalf("HasInFlightRunForTask(%s): %v", tc.status, err)
+			}
+			if got != tc.want {
+				t.Errorf("HasInFlightRunForTask(status=%s) = %v, want %v", tc.status, got, tc.want)
+			}
+		})
+	}
+}
+
+func TestHasInFlightRunForTask_ScopedToTheNamedTask(t *testing.T) {
+	repo := newTestRepo(t)
+	ctx := context.Background()
+
+	seedCancelRun(t, repo, "other-task", "claimed", nil, nil)
+
+	got, err := repo.HasInFlightRunForTask(ctx, "quiet-task")
+	if err != nil {
+		t.Fatalf("HasInFlightRunForTask: %v", err)
+	}
+	if got {
+		t.Error("got true, want false — the only in-flight run belongs to another task")
+	}
+}
+
+func TestHasInFlightRunForTask_EmptyTaskIDIsNotInFlight(t *testing.T) {
+	repo := newTestRepo(t)
+	ctx := context.Background()
+
+	// An empty task ID names no task, so it must never be considered in flight.
+	// HasInFlightRunForTask returns false before the query runs, making this
+	// independent of the dialect; this test pins that early return.
+	seedCancelRun(t, repo, "", "queued", nil, nil)
+
+	got, err := repo.HasInFlightRunForTask(ctx, "")
+	if err != nil {
+		t.Fatalf("HasInFlightRunForTask(\"\"): %v", err)
+	}
+	if got {
+		t.Error("got true, want false — an empty task ID names no task")
+	}
+}
+
+func TestHasInFlightRunForTask_AnyInFlightRunCounts(t *testing.T) {
+	repo := newTestRepo(t)
+	ctx := context.Background()
+
+	// A task that has already run once and is now running again: the terminal
+	// run must not mask the live one.
+	finishedAt := time.Now().UTC().Add(-time.Hour)
+	seedCancelRun(t, repo, "busy-task", "finished", nil, &finishedAt)
+	seedCancelRun(t, repo, "busy-task", "queued", nil, nil)
+
+	got, err := repo.HasInFlightRunForTask(ctx, "busy-task")
+	if err != nil {
+		t.Fatalf("HasInFlightRunForTask: %v", err)
+	}
+	if !got {
+		t.Error("got false, want true — the task has a queued run alongside a finished one")
+	}
+}

apps/backend/internal/orchestrator/idle_session_reaper.go

@@ -157,7 +157,8 @@ func (r *idleSessionReaper) runLoop(ctx context.Context, tick func(ctx context.C
 // startIdleSessionReaper wires the reaper into Service.Start. The
 // reaper captures the canonical fail-closed reclaim primitive so
 // the per-tick work is a thin scan + loop, not a new policy. The tick also
-// drives reclaimStuckSignalSessionsOnce (stuck_signal_watchdog.go) and
+// drives reclaimStuckSignalSessionsOnce (stuck_signal_watchdog.go),
+// detectOfficeDecisionWaitingOnce (office_decision_stall_watchdog.go) and
 // reapStalePendingMovesOnce (pending_move_reaper.go) — further scans
 // sharing this ticker rather than background goroutines of their own,
 // preserving the reaper's single-goroutine-owner invariant. The 30s
@@ -170,6 +171,7 @@ func (s *Service) startIdleSessionReaper(ctx context.Context) {
 	if !s.idleReaper.start(ctx, func(tickCtx context.Context) {
 		s.reclaimIdleSessionsOnce(tickCtx)
 		s.reclaimStuckSignalSessionsOnce(tickCtx)
+		s.detectOfficeDecisionWaitingOnce(tickCtx)
 		s.reapStalePendingMovesOnce(tickCtx)
 	}) {
 		return

apps/backend/internal/orchestrator/office_decision_stall_test.go

@@ -0,0 +1,613 @@
+package orchestrator
+
+import (
+	"context"
+	"errors"
+	"expvar"
+	"fmt"
+	"sync"
+	"testing"
+	"time"
+
+	taskmodels "github.com/kandev/kandev/internal/task/models"
+	sqliterepo "github.com/kandev/kandev/internal/task/repository/sqlite"
+	"github.com/kandev/kandev/internal/workflow/engine"
+	wfmodels "github.com/kandev/kandev/internal/workflow/models"
+	v1 "github.com/kandev/kandev/pkg/api/v1"
+)
+
+// Doubles for the decision-waiting detector. Each can be told to fail, because
+// "the input could not be read" is a behaviour the detector must get right
+// (fail closed, count the reason) and not merely a plumbing accident.
+
+type stallParticipantStore struct {
+	participants []engine.ParticipantInfo
+	err          error
+}
+
+func (f *stallParticipantStore) ListStepParticipants(_ context.Context, _, _ string) ([]engine.ParticipantInfo, error) {
+	return f.participants, f.err
+}
+
+func (f *stallParticipantStore) ListTaskParticipants(_ context.Context, _ string) ([]engine.ParticipantInfo, error) {
+	return f.participants, f.err
+}
+
+type stallDecisionStore struct {
+	decisions []engine.DecisionInfo
+	err       error
+	// recorded and cleared count writes. The detector must never perform
+	// either: recording a decision would forge the quorum the Office gates
+	// exist to require, and clearing one would discard a real reviewer's.
+	recorded int
+	cleared  int
+}
+
+func (f *stallDecisionStore) ListStepDecisions(_ context.Context, _, _ string) ([]engine.DecisionInfo, error) {
+	return f.decisions, f.err
+}
+
+func (f *stallDecisionStore) RecordStepDecision(_ context.Context, _ engine.DecisionInfo) error {
+	f.recorded++
+	return nil
+}
+
+func (f *stallDecisionStore) ClearStepDecisions(_ context.Context, _, _ string) (int64, error) {
+	f.cleared++
+	return 0, nil
+}
+
+// stallRunQueue counts QueueRun calls so "the detector queues no run"
+// (AC-003.2) is an assertion rather than an appeal to reading the source.
+type stallRunQueue struct {
+	calls int
+}
+
+func (q *stallRunQueue) QueueRun(_ context.Context, _ engine.QueueRunRequest) (engine.QueueOutcome, error) {
+	q.calls++
+	return engine.QueueOutcomeQueued, nil
+}
+
+type stallRunReader struct {
+	inFlight bool
+	err      error
+	calls    int
+}
+
+func (f *stallRunReader) HasInFlightRunForTask(_ context.Context, _ string) (bool, error) {
+	f.calls++
+	return f.inFlight, f.err
+}
+
+func decidingSeat() []engine.ParticipantInfo {
+	return []engine.ParticipantInfo{{
+		ID:               "seat-1",
+		StepID:           "step1",
+		Role:             string(wfmodels.ParticipantRoleReviewer),
+		AgentProfileID:   "agent-1",
+		DecisionRequired: true,
+	}}
+}
+
+func officeDecisionWaitingCount() int64 {
+	v := officeStallDecisionWaitingTotal.Get(officeStallLabel("detector", "decision_waiting"))
+	iv, ok := v.(*expvar.Int)
+	if !ok {
+		return 0
+	}
+	return iv.Value()
+}
+
+func officeStallSkipCount(reason string) int64 {
+	v := officeStallDetectorSkippedTotal.Get(officeStallLabel("reason", reason))
+	iv, ok := v.(*expvar.Int)
+	if !ok {
+		return 0
+	}
+	return iv.Value()
+}
+
+// seedDecisionWaitingTask makes an existing task look Office-owned, seats a
+// decision-required reviewer at its step, and backdates the task past the
+// threshold — the three facts the candidate query selects on.
+func seedDecisionWaitingTask(t *testing.T, repo *sqliterepo.Repository, taskID, stepID string) {
+	t.Helper()
+	ctx := context.Background()
+
+	// IsFromOffice is a derived projection, not a stored column, so take the
+	// project_id branch rather than setting the struct field (which the write
+	// path discards). Same reason as seedOfficeStrandedSignal.
+	if _, err := repo.DB().ExecContext(ctx,
+		`UPDATE tasks SET project_id = 'proj1' WHERE id = ?`, taskID); err != nil {
+		t.Fatalf("mark task as office: %v", err)
+	}
+	if _, err := repo.DB().ExecContext(ctx, `
+		INSERT INTO workflow_step_participants
+			(id, step_id, task_id, role, agent_profile_id, decision_required, position)
+		VALUES ('seat-'||?, ?, '', 'reviewer', 'agent-1', 1, 0)
+	`, taskID, stepID); err != nil {
+		t.Fatalf("seed decision seat: %v", err)
+	}
+	quiet := time.Now().UTC().Add(-2 * officeDecisionWaitingThreshold)
+	if _, err := repo.DB().ExecContext(ctx,
+		`UPDATE tasks SET updated_at = ? WHERE id = ?`, quiet, taskID); err != nil {
+		t.Fatalf("backdate task: %v", err)
+	}
+
+	task, err := repo.GetTask(ctx, taskID)
+	if err != nil {
+		t.Fatalf("get task: %v", err)
+	}
+	if !task.IsFromOffice {
+		t.Fatalf("precondition: task %s should read as an Office task", taskID)
+	}
+}
+
+// decisionWaitingService builds a Service with the detector's three
+// dependencies wired to the supplied doubles.
+func decisionWaitingService(
+	t *testing.T, repo *sqliterepo.Repository,
+	participants *stallParticipantStore, decisions *stallDecisionStore, runs *stallRunReader,
+) *Service {
+	t.Helper()
+	svc := createTestService(repo, officeStallStepGetter(), newMockTaskRepo())
+	if participants != nil {
+		svc.SetEngineParticipantStore(participants)
+	}
+	if decisions != nil {
+		svc.SetEngineDecisionStore(decisions)
+	}
+	if runs != nil {
+		svc.SetOfficeRunInFlightReader(runs)
+	}
+	return svc
+}
+
+// AC-002.1: an Office task past the threshold with an undecided seat and no
+// run in flight is surfaced.
+func TestDetectOfficeDecisionWaitingOnce_SurfacesWaitingTask(t *testing.T) {
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	seedDecisionWaitingTask(t, repo, "t1", "step1")
+	runs := &stallRunReader{}
+	decisions := &stallDecisionStore{}
+	queue := &stallRunQueue{}
+	svc := decisionWaitingService(t, repo,
+		&stallParticipantStore{participants: decidingSeat()}, decisions, runs)
+	svc.SetEngineRunQueue(queue)
+
+	before := officeDecisionWaitingCount()
+	svc.detectOfficeDecisionWaitingOnce(context.Background())
+
+	if got := officeDecisionWaitingCount() - before; got != 1 {
+		t.Fatalf("decision-waiting counter delta = %d, want 1", got)
+	}
+	if runs.calls != 1 {
+		t.Fatalf("run reader called %d times, want 1", runs.calls)
+	}
+	assertOfficeDecisionNotActedOn(t, repo, "t1", "step1", decisions, queue)
+}
+
+// A bounded page must not make the oldest candidates a permanent exclusion
+// list. Every candidate past the first page still needs one evaluation during
+// a scan pass.
+func TestDetectOfficeDecisionWaitingOnce_ProcessesCandidatesBeyondFirstPage(t *testing.T) {
+	const candidateCount = 201
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "task-000", "s-seed", "step1")
+	seedDecisionWaitingTask(t, repo, "task-000", "step1")
+	for i := 1; i < candidateCount; i++ {
+		taskID := fmt.Sprintf("task-%03d", i)
+		now := time.Now().UTC()
+		if err := repo.CreateTask(context.Background(), &taskmodels.Task{
+			ID:             taskID,
+			WorkspaceID:    "ws1",
+			WorkflowID:     "wf1",
+			WorkflowStepID: "step1",
+			Title:          taskID,
+			State:          v1.TaskStateInProgress,
+			CreatedAt:      now,
+			UpdatedAt:      now,
+		}); err != nil {
+			t.Fatalf("create task %s: %v", taskID, err)
+		}
+		seedDecisionWaitingTask(t, repo, taskID, "step1")
+	}
+	svc := decisionWaitingService(t, repo,
+		&stallParticipantStore{participants: decidingSeat()}, &stallDecisionStore{},
+		&stallRunReader{})
+
+	before := officeDecisionWaitingCount()
+	svc.detectOfficeDecisionWaitingOnce(context.Background())
+
+	if got := officeDecisionWaitingCount() - before; got != candidateCount {
+		t.Fatalf("decision-waiting counter delta = %d, want %d", got, candidateCount)
+	}
+	if _, seen := svc.officeDecisionWaiting.Load(officeDecisionWaitingKey("task-200", "step1")); !seen {
+		t.Fatal("candidate beyond the first page was not evaluated")
+	}
+}
+
+// The Office dependency readers are wired after Service.Start. The detector
+// can run during that interval, so late wiring and detector reads must be race
+// free.
+func TestOfficeDecisionWaitingLateRunReaderWiringIsRaceFree(t *testing.T) {
+	repo := setupTestRepo(t)
+	seedTaskWithoutSession(t, repo, "t1", "step1")
+	seedDecisionWaitingTask(t, repo, "t1", "step1")
+	svc := decisionWaitingService(t, repo,
+		&stallParticipantStore{participants: decidingSeat()}, &stallDecisionStore{},
+		&stallRunReader{})
+	ctx := context.Background()
+
+	const iterations = 1000
+	start := make(chan struct{})
+	var wg sync.WaitGroup
+	wg.Add(2)
+	go func() {
+		defer wg.Done()
+		<-start
+		for i := 0; i < iterations; i++ {
+			svc.detectOfficeDecisionWaitingOnce(ctx)
+		}
+	}()
+	go func() {
+		defer wg.Done()
+		close(start)
+		for i := 0; i < iterations; i++ {
+			svc.SetOfficeRunInFlightReader(&stallRunReader{inFlight: i%2 == 0})
+		}
+	}()
+	wg.Wait()
+}
+
+// AC-002.2: the false-positive guard. A task with a claimed run is being
+// worked on, so it must not be reported however long its seat has gone
+// undecided. This is the case named in the definition of done.
+func TestDetectOfficeDecisionWaitingOnce_ClaimedRunIsNotSurfaced(t *testing.T) {
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	seedDecisionWaitingTask(t, repo, "t1", "step1")
+	svc := decisionWaitingService(t, repo,
+		&stallParticipantStore{participants: decidingSeat()}, &stallDecisionStore{},
+		&stallRunReader{inFlight: true})
+
+	before := officeDecisionWaitingCount()
+	svc.detectOfficeDecisionWaitingOnce(context.Background())
+
+	if got := officeDecisionWaitingCount() - before; got != 0 {
+		t.Fatalf("counter delta = %d, want 0 — a task with a run in flight is healthy", got)
+	}
+}
+
+// AC-002.3: a decision already recorded for the current step suppresses
+// surfacing.
+func TestDetectOfficeDecisionWaitingOnce_RecordedDecisionSuppresses(t *testing.T) {
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	seedDecisionWaitingTask(t, repo, "t1", "step1")
+	runs := &stallRunReader{}
+	svc := decisionWaitingService(t, repo,
+		&stallParticipantStore{participants: decidingSeat()},
+		&stallDecisionStore{decisions: []engine.DecisionInfo{{
+			ID: "d1", TaskID: "t1", StepID: "step1", Decision: "approve",
+		}}},
+		runs)
+
+	before := officeDecisionWaitingCount()
+	svc.detectOfficeDecisionWaitingOnce(context.Background())
+
+	if got := officeDecisionWaitingCount() - before; got != 0 {
+		t.Fatalf("counter delta = %d, want 0 — the step already has a decision", got)
+	}
+	if runs.calls != 0 {
+		t.Fatalf("run reader called %d times; the decision check must reject first", runs.calls)
+	}
+}
+
+// AC-002.3 (superseded case): ListStepDecisions returns superseded rows
+// alongside active ones (its own documented contract), so a step whose only
+// decision is superseded must still read as undecided. maybeSupersedeOnRework
+// leaves the row in place with superseded_at set rather than deleting it, so
+// treating any non-empty result as "decided" would permanently suppress the
+// repeat stall this detector exists to find.
+func TestDetectOfficeDecisionWaitingOnce_SupersededDecisionDoesNotSuppress(t *testing.T) {
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	seedDecisionWaitingTask(t, repo, "t1", "step1")
+	runs := &stallRunReader{}
+	supersededAt := time.Now().UTC().Add(-time.Hour)
+	svc := decisionWaitingService(t, repo,
+		&stallParticipantStore{participants: decidingSeat()},
+		&stallDecisionStore{decisions: []engine.DecisionInfo{{
+			ID: "d1", TaskID: "t1", StepID: "step1", Decision: "approve",
+			SupersededAt: &supersededAt,
+		}}},
+		runs)
+
+	before := officeDecisionWaitingCount()
+	svc.detectOfficeDecisionWaitingOnce(context.Background())
+
+	if got := officeDecisionWaitingCount() - before; got != 1 {
+		t.Fatalf("counter delta = %d, want 1 — a superseded-only decision set is undecided", got)
+	}
+	if runs.calls != 1 {
+		t.Fatalf("run reader called %d times, want 1 — the superseded row must not short-circuit the check", runs.calls)
+	}
+}
+
+// AC-002.4: seats that owe no decision cannot strand a task.
+func TestDetectOfficeDecisionWaitingOnce_NonDecidingSeatsAreNotSurfaced(t *testing.T) {
+	cases := []struct {
+		name string
+		seat engine.ParticipantInfo
+	}{
+		{"watcher", engine.ParticipantInfo{Role: string(wfmodels.ParticipantRoleWatcher), DecisionRequired: true}},
+		{"collaborator", engine.ParticipantInfo{Role: string(wfmodels.ParticipantRoleCollaborator), DecisionRequired: true}},
+		{"runner", engine.ParticipantInfo{Role: string(wfmodels.ParticipantRoleRunner), DecisionRequired: true}},
+		{"reviewer not required", engine.ParticipantInfo{Role: string(wfmodels.ParticipantRoleReviewer), DecisionRequired: false}},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			repo := setupTestRepo(t)
+			seedSession(t, repo, "t1", "s1", "step1")
+			seedDecisionWaitingTask(t, repo, "t1", "step1")
+			runs := &stallRunReader{}
+			svc := decisionWaitingService(t, repo,
+				&stallParticipantStore{participants: []engine.ParticipantInfo{tc.seat}},
+				&stallDecisionStore{}, runs)
+
+			before := officeDecisionWaitingCount()
+			svc.detectOfficeDecisionWaitingOnce(context.Background())
+
+			if got := officeDecisionWaitingCount() - before; got != 0 {
+				t.Fatalf("counter delta = %d, want 0 — a %s seat owes no decision", got, tc.name)
+			}
+			if runs.calls != 0 {
+				t.Fatalf("run reader called %d times; the seat check must reject first", runs.calls)
+			}
+		})
+	}
+}
+
+// AC-002.5: every unreadable input fails closed and is counted, so a silently
+// degraded detector is distinguishable from a quiet system.
+func TestDetectOfficeDecisionWaitingOnce_UnreadableInputsFailClosed(t *testing.T) {
+	readErr := errors.New("boom")
+	cases := []struct {
+		name         string
+		participants *stallParticipantStore
+		decisions    *stallDecisionStore
+		runs         *stallRunReader
+		reason       string
+	}{
+		{
+			name:   "participant store unwired",
+			runs:   &stallRunReader{},
+			reason: officeStallSkipParticipantStore,
+		},
+		{
+			name:         "decision store unwired",
+			participants: &stallParticipantStore{participants: decidingSeat()},
+			runs:         &stallRunReader{},
+			reason:       officeStallSkipDecisionStore,
+		},
+		{
+			name:         "participant read failed",
+			participants: &stallParticipantStore{err: readErr},
+			decisions:    &stallDecisionStore{},
+			runs:         &stallRunReader{},
+			reason:       officeStallSkipParticipantReadFailed,
+		},
+		{
+			name:         "decision read failed",
+			participants: &stallParticipantStore{participants: decidingSeat()},
+			decisions:    &stallDecisionStore{err: readErr},
+			runs:         &stallRunReader{},
+			reason:       officeStallSkipDecisionReadFailed,
+		},
+		{
+			name:         "run reader unwired",
+			participants: &stallParticipantStore{participants: decidingSeat()},
+			decisions:    &stallDecisionStore{},
+			reason:       officeStallSkipRunReaderUnwired,
+		},
+		{
+			name:         "run reader errored",
+			participants: &stallParticipantStore{participants: decidingSeat()},
+			decisions:    &stallDecisionStore{},
+			runs:         &stallRunReader{err: readErr},
+			reason:       officeStallSkipRunReaderError,
+		},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			repo := setupTestRepo(t)
+			seedSession(t, repo, "t1", "s1", "step1")
+			seedDecisionWaitingTask(t, repo, "t1", "step1")
+			svc := decisionWaitingService(t, repo, tc.participants, tc.decisions, tc.runs)
+
+			beforeSurfaced := officeDecisionWaitingCount()
+			beforeSkipped := officeStallSkipCount(tc.reason)
+			svc.detectOfficeDecisionWaitingOnce(context.Background())
+
+			if got := officeDecisionWaitingCount() - beforeSurfaced; got != 0 {
+				t.Errorf("counter delta = %d, want 0 — an unreadable input must not surface", got)
+			}
+			if got := officeStallSkipCount(tc.reason) - beforeSkipped; got != 1 {
+				t.Errorf("skip counter %q delta = %d, want 1", tc.reason, got)
+			}
+		})
+	}
+}
+
+// AC-002.6: a non-Office task is never evaluated. The exclusion here is the
+// mirror image of the stranded-signal path's: there, Office must not be
+// excluded from detection; here, non-Office must not be included.
+func TestDetectOfficeDecisionWaitingOnce_NonOfficeTaskIsNeverEvaluated(t *testing.T) {
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	ctx := context.Background()
+	// Everything a candidate needs except Office ownership.
+	if _, err := repo.DB().ExecContext(ctx, `
+		INSERT INTO workflow_step_participants
+			(id, step_id, task_id, role, agent_profile_id, decision_required, position)
+		VALUES ('seat-kanban', 'step1', '', 'reviewer', 'agent-1', 1, 0)
+	`); err != nil {
+		t.Fatalf("seed decision seat: %v", err)
+	}
+	quiet := time.Now().UTC().Add(-2 * officeDecisionWaitingThreshold)
+	if _, err := repo.DB().ExecContext(ctx,
+		`UPDATE tasks SET updated_at = ? WHERE id = 't1'`, quiet); err != nil {
+		t.Fatalf("backdate task: %v", err)
+	}
+	task, err := repo.GetTask(ctx, "t1")
+	if err != nil {
+		t.Fatalf("get task: %v", err)
+	}
+	if task.IsFromOffice {
+		t.Fatalf("precondition: t1 must read as a kanban task")
+	}
+
+	runs := &stallRunReader{}
+	svc := decisionWaitingService(t, repo,
+		&stallParticipantStore{participants: decidingSeat()}, &stallDecisionStore{}, runs)
+
+	before := officeDecisionWaitingCount()
+	svc.detectOfficeDecisionWaitingOnce(ctx)
+
+	if got := officeDecisionWaitingCount() - before; got != 0 {
+		t.Fatalf("counter delta = %d, want 0 — a kanban task is out of scope", got)
+	}
+	if runs.calls != 0 {
+		t.Fatalf("run reader called %d times for a kanban task", runs.calls)
+	}
+}
+
+// A still-waiting task is reported once, not on every 30-second tick.
+func TestDetectOfficeDecisionWaitingOnce_ReportsOncePerTaskAndStep(t *testing.T) {
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	seedDecisionWaitingTask(t, repo, "t1", "step1")
+	svc := decisionWaitingService(t, repo,
+		&stallParticipantStore{participants: decidingSeat()}, &stallDecisionStore{},
+		&stallRunReader{})
+
+	before := officeDecisionWaitingCount()
+	for range 3 {
+		svc.detectOfficeDecisionWaitingOnce(context.Background())
+	}
+	if got := officeDecisionWaitingCount() - before; got != 1 {
+		t.Fatalf("counter delta across three scans = %d, want 1", got)
+	}
+}
+
+// A task that stops being a candidate is pruned from the dedupe map, so if it
+// stalls again later it is reported again rather than silently swallowed.
+func TestDetectOfficeDecisionWaitingOnce_PrunesResolvedTasks(t *testing.T) {
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	seedDecisionWaitingTask(t, repo, "t1", "step1")
+	svc := decisionWaitingService(t, repo,
+		&stallParticipantStore{participants: decidingSeat()}, &stallDecisionStore{},
+		&stallRunReader{})
+	ctx := context.Background()
+
+	before := officeDecisionWaitingCount()
+	svc.detectOfficeDecisionWaitingOnce(ctx)
+
+	// The reviewer finally acts: the task row is touched, so it is no longer
+	// quiet and drops out of the candidate set.
+	if _, err := repo.DB().ExecContext(ctx,
+		`UPDATE tasks SET updated_at = ? WHERE id = 't1'`, time.Now().UTC()); err != nil {
+		t.Fatalf("touch task: %v", err)
+	}
+	svc.detectOfficeDecisionWaitingOnce(ctx)
+	if _, seen := svc.officeDecisionWaiting.Load(officeDecisionWaitingKey("t1", "step1")); seen {
+		t.Fatal("dedupe entry survived a scan in which the task was no longer a candidate")
+	}
+
+	// It stalls again: a new report, not a swallowed repeat.
+	quiet := time.Now().UTC().Add(-2 * officeDecisionWaitingThreshold)
+	if _, err := repo.DB().ExecContext(ctx,
+		`UPDATE tasks SET updated_at = ? WHERE id = 't1'`, quiet); err != nil {
+		t.Fatalf("re-backdate task: %v", err)
+	}
+	svc.detectOfficeDecisionWaitingOnce(ctx)
+	if got := officeDecisionWaitingCount() - before; got != 2 {
+		t.Fatalf("counter delta = %d, want 2 — a fresh stall is a fresh report", got)
+	}
+}
+
+// A task whose decision is recorded, then superseded by a fresh requirement,
+// must be reported again even though tasks.updated_at never moves and the
+// task never leaves the cheap SQL candidate window: recording a decision
+// must clear the dedupe key, not merely suppress the tick it was recorded on
+// (regression for the "live" set tracking the SQL prefilter instead of the
+// full predicate).
+func TestDetectOfficeDecisionWaitingOnce_RestallAfterDecisionIsReportedAgain(t *testing.T) {
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	seedDecisionWaitingTask(t, repo, "t1", "step1")
+	decisions := &stallDecisionStore{}
+	svc := decisionWaitingService(t, repo,
+		&stallParticipantStore{participants: decidingSeat()}, decisions,
+		&stallRunReader{})
+	ctx := context.Background()
+
+	before := officeDecisionWaitingCount()
+	svc.detectOfficeDecisionWaitingOnce(ctx)
+	if got := officeDecisionWaitingCount() - before; got != 1 {
+		t.Fatalf("first scan counter delta = %d, want 1", got)
+	}
+
+	// The reviewer decides. tasks.updated_at is untouched, so the task stays
+	// in the cheap SQL candidate window even though the full predicate no
+	// longer holds.
+	decisions.decisions = []engine.DecisionInfo{{
+		ID: "d1", TaskID: "t1", StepID: "step1", Decision: "reject",
+	}}
+	svc.detectOfficeDecisionWaitingOnce(ctx)
+	if _, seen := svc.officeDecisionWaiting.Load(officeDecisionWaitingKey("t1", "step1")); seen {
+		t.Fatal("dedupe entry survived a scan in which the decision was recorded")
+	}
+
+	// Rework: a fresh requirement supersedes the old decision, still at the
+	// same step, still with the same stale tasks.updated_at.
+	supersededAt := time.Now().UTC()
+	decisions.decisions = []engine.DecisionInfo{{
+		ID: "d1", TaskID: "t1", StepID: "step1", Decision: "reject",
+		SupersededAt: &supersededAt,
+	}}
+	svc.detectOfficeDecisionWaitingOnce(ctx)
+	if got := officeDecisionWaitingCount() - before; got != 2 {
+		t.Fatalf("counter delta after rework = %d, want 2 — the rework stall must be reported again", got)
+	}
+}
+
+// AC-003.1 through AC-003.3: surfacing is the entire action. No decision row,
+// no queued run, no step transition — forging any of those would manufacture
+// the quorum the Office gates exist to require.
+func assertOfficeDecisionNotActedOn(
+	t *testing.T, repo *sqliterepo.Repository, taskID, stepID string,
+	decisions *stallDecisionStore, queue *stallRunQueue,
+) {
+	t.Helper()
+
+	task, err := repo.GetTask(context.Background(), taskID)
+	if err != nil {
+		t.Fatalf("get task: %v", err)
+	}
+	if task.WorkflowStepID != stepID {
+		t.Fatalf("workflow step moved to %q; the detector applied a transition", task.WorkflowStepID)
+	}
+	if decisions.recorded != 0 {
+		t.Fatalf("detector recorded %d decisions; it must never manufacture a quorum", decisions.recorded)
+	}
+	if decisions.cleared != 0 {
+		t.Fatalf("detector cleared decisions %d times; it must never discard a real one", decisions.cleared)
+	}
+	if queue.calls != 0 {
+		t.Fatalf("detector queued %d runs; surfacing is the whole action", queue.calls)
+	}
+}

apps/backend/internal/orchestrator/office_decision_stall_watchdog.go

@@ -0,0 +1,288 @@
+package orchestrator
+
+import (
+	"context"
+	"time"
+
+	"go.uber.org/zap"
+
+	"github.com/kandev/kandev/internal/task/models"
+	"github.com/kandev/kandev/internal/workflow/engine"
+	wfmodels "github.com/kandev/kandev/internal/workflow/models"
+)
+
+// Office decision-waiting detector.
+//
+// The stuck-signal watchdog next door only sees a task whose session accepted
+// a completion signal. It cannot see the other Office stall: a task parked at
+// a review or approval step where the seat that owes a decision never records
+// one. Nothing is running, nothing failed, no signal is pending — the card
+// simply stops moving, and the only present symptom is the absence of an
+// event. This pass detects that state and reports it.
+//
+// It never acts on what it finds (REQ-OFFICE-STALL-VISIBILITY-003). Advancing
+// the task would mean manufacturing the decision the quorum gate is there to
+// require, so this file contains no write to workflow_step_decisions, no
+// queued run, and no step transition — only a log line and a counter. That
+// restraint is the feature, not a limitation of it.
+const (
+	// officeDecisionWaitingThreshold is how long an Office task may sit at a
+	// step with an undecided decision-required seat before the detector
+	// reports it.
+	//
+	// PROVISIONAL. 60 minutes is a first guess, not a measurement: a human
+	// reviewer who steps away for ten minutes is behaving normally, and an
+	// agent reviewer that has genuinely died is equally stuck at ten minutes
+	// or at sixty, so the cost of waiting is only reporting latency while the
+	// cost of being early is a log full of healthy tasks. Re-tune once the
+	// counter has produced real distribution data.
+	//
+	// Deliberately NOT stuckSignalWatchdogThreshold. The two constants
+	// measure unrelated phenomena — one an agent process that stopped
+	// emitting events mid-turn, the other a decision nobody has made yet —
+	// and sharing the value would make a future tuning of session silence
+	// silently retune this, in the wrong direction and with no test to catch
+	// it.
+	officeDecisionWaitingThreshold = 60 * time.Minute
+)
+
+// officeDecisionCandidateLister is implemented by the durable task repository.
+// A narrow optional interface, mirroring stuckSignalSessionLister, so test
+// doubles that do not exercise this scan stay compatible.
+type officeDecisionCandidateLister interface {
+	ListOfficeDecisionWaitCandidates(
+		ctx context.Context,
+		quietSince time.Time,
+		after *models.OfficeDecisionWaitCursor,
+	) ([]models.OfficeDecisionWaitCandidate, *models.OfficeDecisionWaitCursor, error)
+}
+
+// officeRunInFlightReader answers whether a task still has a queued or claimed
+// run. It is the detector's false-positive guard: a task with work in flight
+// is being worked on, not stalled, however long its seat has gone undecided.
+// Implemented by the Office SQLite repository (HasInFlightRunForTask) and
+// wired from internal/backendapp, where that repository is constructed; the
+// orchestrator does not otherwise reach the runs table, so this stays an
+// injected dependency rather than a repo cast.
+type officeRunInFlightReader interface {
+	HasInFlightRunForTask(ctx context.Context, taskID string) (bool, error)
+}
+
+// SetOfficeRunInFlightReader wires the run-queue read the Office
+// decision-waiting detector needs. Until it is wired the detector reports
+// nothing and counts every evaluation as skipped, rather than reporting
+// tasks it cannot prove are idle.
+func (s *Service) SetOfficeRunInFlightReader(reader officeRunInFlightReader) {
+	s.officeStallDependenciesMu.Lock()
+	s.officeRunInFlight = reader
+	s.officeStallDependenciesMu.Unlock()
+}
+
+// officeDecisionWaitingKey identifies one observation for dedupe purposes.
+// Keyed by task and step so a task that moves to another decision step — or
+// returns to this one after a rejection round — is reported again, while a
+// task sitting still is reported once.
+func officeDecisionWaitingKey(taskID, stepID string) string {
+	return taskID + "\x00" + stepID
+}
+
+// detectOfficeDecisionWaitingOnce is the per-tick scan, called from the
+// idle-session reaper's existing ticker alongside reclaimStuckSignalSessionsOnce
+// rather than owning a second background goroutine. It walks bounded keyset
+// pages until the scan budget expires or the candidate set is complete.
+func (s *Service) detectOfficeDecisionWaitingOnce(ctx context.Context) {
+	lister, ok := s.repo.(officeDecisionCandidateLister)
+	if !ok {
+		officeStallSkipped(officeStallSkipCandidateListUnwired)
+		return
+	}
+	now := time.Now().UTC()
+	scanCtx, cancel := context.WithTimeout(ctx, stuckSignalScanBudget)
+	defer cancel()
+	live := make(map[string]struct{})
+	var cursor *models.OfficeDecisionWaitCursor
+	scanned := 0
+	for {
+		if scanCtx.Err() != nil {
+			s.logOfficeDecisionScanBudgetExceeded(scanned)
+			return
+		}
+		candidates, next, err := lister.ListOfficeDecisionWaitCandidates(
+			scanCtx, now.Add(-officeDecisionWaitingThreshold), cursor,
+		)
+		if err != nil {
+			if scanCtx.Err() != nil {
+				s.logOfficeDecisionScanBudgetExceeded(scanned)
+				return
+			}
+			officeStallSkipped(officeStallSkipCandidateListFailed)
+			s.logger.Warn("office stall watchdog: listing decision-wait candidates failed; tick skipped",
+				zap.Error(err))
+			return
+		}
+		for _, candidate := range candidates {
+			if scanCtx.Err() != nil {
+				s.logOfficeDecisionScanBudgetExceeded(scanned)
+				return
+			}
+			scanned++
+			if s.evaluateOfficeDecisionWaiting(scanCtx, candidate, now) {
+				live[officeDecisionWaitingKey(candidate.TaskID, candidate.StepID)] = struct{}{}
+			}
+		}
+		if next == nil || len(candidates) == 0 {
+			break
+		}
+		cursor = next
+	}
+	s.pruneOfficeDecisionWaiting(live)
+}
+
+func (s *Service) logOfficeDecisionScanBudgetExceeded(scanned int) {
+	s.logger.Warn("office stall watchdog: decision-wait scan budget exceeded; deferring remaining candidates to the next tick",
+		zap.Int("candidates_scanned", scanned))
+}
+
+// evaluateOfficeDecisionWaiting applies the remaining predicate to one
+// candidate, reports it if every clause holds, and reports whether the
+// candidate should stay live for pruning purposes. The order is deliberate:
+// the cheapest, most-selective reads run first and the runs-table read runs
+// last, because most candidates are rejected before reaching it.
+//
+// Live must track the full predicate, not just the cheap SQL prefilter that
+// produced the candidate: the query has no way to see a recorded decision or
+// an in-flight run, so a resolved task keeps matching it for as long as
+// tasks.updated_at stays quiet. Marking it live from the SQL match alone
+// would keep its dedupe key past resolution, and a rework round that
+// re-stalls the same task at the same step — without ever touching
+// updated_at — would then find the old key already present and be silently
+// swallowed by surfaceOfficeDecisionWaiting's LoadOrStore, contradicting
+// officeDecisionWaitingKey's own "returns to this one after a rejection
+// round — is reported again" contract.
+//
+// Every unreadable input fails closed to "do not report" and is counted. A
+// detector that guessed on missing input would report healthy tasks, and the
+// only thing this feature produces is reports. Failing closed here also
+// means an unreadable input is not live, so a transient read failure costs at
+// most one duplicate report on a later successful tick rather than risking a
+// permanently swallowed one.
+func (s *Service) evaluateOfficeDecisionWaiting(
+	ctx context.Context, candidate models.OfficeDecisionWaitCandidate, now time.Time,
+) bool {
+	if !s.officeStepAwaitsDecision(ctx, candidate) {
+		return false
+	}
+	_, _, reader := s.officeStallDependencies()
+	if reader == nil {
+		officeStallSkipped(officeStallSkipRunReaderUnwired)
+		return false
+	}
+	inFlight, err := reader.HasInFlightRunForTask(ctx, candidate.TaskID)
+	if err != nil {
+		officeStallSkipped(officeStallSkipRunReaderError)
+		return false
+	}
+	if inFlight {
+		return false
+	}
+	s.surfaceOfficeDecisionWaiting(candidate, now)
+	return true
+}
+
+// officeStepAwaitsDecision reports whether the candidate's current step holds
+// a decision-required reviewer or approver seat with no decision recorded
+// against it. Both stores are required: without them the detector cannot tell
+// a genuinely undecided step from one it simply could not read.
+func (s *Service) officeStepAwaitsDecision(
+	ctx context.Context, candidate models.OfficeDecisionWaitCandidate,
+) bool {
+	participantsStore, decisionsStore, _ := s.officeStallDependencies()
+	if participantsStore == nil {
+		officeStallSkipped(officeStallSkipParticipantStore)
+		return false
+	}
+	if decisionsStore == nil {
+		officeStallSkipped(officeStallSkipDecisionStore)
+		return false
+	}
+	participants, err := participantsStore.ListStepParticipants(ctx, candidate.StepID, candidate.TaskID)
+	if err != nil {
+		officeStallSkipped(officeStallSkipParticipantReadFailed)
+		return false
+	}
+	if !hasDecidingSeat(participants) {
+		return false
+	}
+	decisions, err := decisionsStore.ListStepDecisions(ctx, candidate.TaskID, candidate.StepID)
+	if err != nil {
+		officeStallSkipped(officeStallSkipDecisionReadFailed)
+		return false
+	}
+	return !hasActiveDecision(decisions)
+}
+
+// hasActiveDecision reports whether any row is not superseded.
+// ListStepDecisions returns superseded rows alongside active ones (see
+// engine.DecisionInfo), and a rework round leaves the superseded row in
+// place rather than deleting it, so a re-entered step must be judged on
+// active rows only or it reads as permanently decided.
+func hasActiveDecision(decisions []engine.DecisionInfo) bool {
+	for _, d := range decisions {
+		if d.SupersededAt == nil {
+			return true
+		}
+	}
+	return false
+}
+
+// hasDecidingSeat reports whether any seat at the step owes a decision.
+// Watcher, collaborator and runner seats never do, and a reviewer seat with
+// DecisionRequired false is an observer — none of them can strand a task, so
+// none of them makes it a candidate.
+func hasDecidingSeat(participants []engine.ParticipantInfo) bool {
+	for _, p := range participants {
+		if !p.DecisionRequired {
+			continue
+		}
+		switch wfmodels.ParticipantRole(p.Role) {
+		case wfmodels.ParticipantRoleReviewer, wfmodels.ParticipantRoleApprover:
+			return true
+		}
+	}
+	return false
+}
+
+// surfaceOfficeDecisionWaiting reports one decision-waiting task exactly once
+// per (task, step). Reporting only — see this file's header for why there is
+// no recovery here.
+func (s *Service) surfaceOfficeDecisionWaiting(
+	candidate models.OfficeDecisionWaitCandidate, now time.Time,
+) {
+	key := officeDecisionWaitingKey(candidate.TaskID, candidate.StepID)
+	if _, seen := s.officeDecisionWaiting.LoadOrStore(key, struct{}{}); seen {
+		return
+	}
+	officeStallDecisionWaitingTotal.Add(officeStallLabel("detector", "decision_waiting"), 1)
+	s.logger.Warn("office stall watchdog: Office task is waiting on a decision that has not been recorded",
+		zap.String("task_id", candidate.TaskID),
+		zap.String("step_id", candidate.StepID),
+		zap.Time("task_updated_at", candidate.UpdatedAt),
+		zap.Duration("quiet_for", now.Sub(candidate.UpdatedAt)),
+		zap.Duration("threshold", officeDecisionWaitingThreshold))
+}
+
+// pruneOfficeDecisionWaiting drops dedupe entries for tasks that are no longer
+// candidates, so the map stays bounded by the number of currently-waiting
+// tasks rather than growing for the process's lifetime. A task that becomes a
+// candidate again after being pruned is correctly reported again: it is a new
+// stall, not a repeat of the old one.
+func (s *Service) pruneOfficeDecisionWaiting(live map[string]struct{}) {
+	s.officeDecisionWaiting.Range(func(key, _ any) bool {
+		if k, ok := key.(string); ok {
+			if _, still := live[k]; !still {
+				s.officeDecisionWaiting.Delete(k)
+			}
+		}
+		return true
+	})
+}

apps/backend/internal/orchestrator/office_stall_metrics.go

@@ -0,0 +1,61 @@
+package orchestrator
+
+import (
+	"expvar"
+	"strings"
+)
+
+// expvar maps for the Office stall detectors, exposed via stdlib's
+// /debug/vars handler. The label model is "key=value;key=value..." so a
+// Prometheus translation layer can split on `;` and `=` later, mirroring
+// internal/office/scheduler/metrics_vars.go.
+//
+// Counters only. A consumer that wants a snapshot of what is currently stuck
+// re-runs the detectors' own predicates against the database; these record
+// events.
+var (
+	// officeStallStrandedSignalTotal counts Office tasks observed holding a
+	// stranded step-completion signal. Labelled by `gate` so it is visible
+	// which of the two watchdog sites observed it.
+	officeStallStrandedSignalTotal = expvar.NewMap("office_stall_stranded_signal_total")
+
+	// officeStallDecisionWaitingTotal counts Office tasks observed parked at
+	// a decision-required step with no run in flight.
+	officeStallDecisionWaitingTotal = expvar.NewMap("office_stall_decision_waiting_total")
+
+	// officeStallDetectorSkippedTotal counts evaluations abandoned because an
+	// input could not be read. Both detectors fail closed, so without this a
+	// silently degraded detector is indistinguishable from a quiet system —
+	// which is the failure mode that matters for a surface-only capability.
+	officeStallDetectorSkippedTotal = expvar.NewMap("office_stall_detector_skipped_total")
+)
+
+// Skip reasons for officeStallDetectorSkippedTotal.
+const (
+	officeStallSkipRunReaderUnwired      = "run_reader_unwired"
+	officeStallSkipRunReaderError        = "run_reader_error"
+	officeStallSkipParticipantStore      = "participant_store_unwired"
+	officeStallSkipDecisionStore         = "decision_store_unwired"
+	officeStallSkipParticipantReadFailed = "participant_read_failed"
+	officeStallSkipDecisionReadFailed    = "decision_read_failed"
+	officeStallSkipCandidateListUnwired  = "candidate_lister_unwired"
+	officeStallSkipCandidateListFailed   = "candidate_list_failed"
+)
+
+// officeStallLabel builds a "k1=v1;k2=v2;..." expvar map key. Returns an
+// empty label for an odd number of arguments rather than guessing.
+func officeStallLabel(pairs ...string) string {
+	if len(pairs)%2 != 0 {
+		return ""
+	}
+	parts := make([]string, 0, len(pairs)/2)
+	for i := 0; i < len(pairs); i += 2 {
+		parts = append(parts, pairs[i]+"="+pairs[i+1])
+	}
+	return strings.Join(parts, ";")
+}
+
+// officeStallSkipped records a fail-closed skip from either detector.
+func officeStallSkipped(reason string) {
+	officeStallDetectorSkippedTotal.Add(officeStallLabel("reason", reason), 1)
+}

apps/backend/internal/orchestrator/office_stall_visibility_test.go

@@ -0,0 +1,328 @@
+package orchestrator
+
+import (
+	"context"
+	"expvar"
+	"testing"
+	"time"
+
+	sqliterepo "github.com/kandev/kandev/internal/task/repository/sqlite"
+
+	"github.com/kandev/kandev/internal/task/models"
+	wfmodels "github.com/kandev/kandev/internal/workflow/models"
+)
+
+// officeStallStepGetter builds a two-step workflow whose first step would
+// auto-advance on turn completion. If the Office exclusion ever regressed, the
+// task would move to step2 and the not-acted-on assertions would fail.
+func officeStallStepGetter() *mockStepGetter {
+	stepGetter := newMockStepGetter()
+	stepGetter.steps["step1"] = &wfmodels.WorkflowStep{
+		ID: "step1", WorkflowID: "wf1", Name: "Step 1", Position: 0,
+		AutoAdvanceRequiresSignal: true,
+		Events: wfmodels.StepEvents{
+			OnTurnComplete: []wfmodels.OnTurnCompleteAction{
+				{Type: wfmodels.OnTurnCompleteMoveToNext},
+			},
+		},
+	}
+	stepGetter.steps["step2"] = &wfmodels.WorkflowStep{
+		ID: "step2", WorkflowID: "wf1", Name: "Step 2", Position: 1,
+	}
+	return stepGetter
+}
+
+// The Office exclusion at both watchdog gates is load-bearing: passing the
+// reconcileWaitingStuckSignalSessionIfDue gate falls through to
+// reconcileStepCompletionSignalLocked, and passing the stuckSignalCandidate
+// gate leads to reclaimStuckSignalSessionOwned. Either one applies the step
+// transition the agent asked for, which for an Office task would record it
+// without the decisions its quorum gate requires. These tests pin the
+// disposition that keeps that from happening.
+func TestStuckSignalDispositionFor(t *testing.T) {
+	const stepID = "step1"
+	signal := models.PendingStepCompletionSignal{StepID: stepID}
+
+	tests := []struct {
+		name    string
+		task    *models.Task
+		session *models.TaskSession
+		want    stuckSignalDisposition
+	}{
+		{
+			name:    "ordinary task at the signalled step is reclaimable",
+			task:    &models.Task{ID: "t1", WorkflowStepID: stepID},
+			session: &models.TaskSession{ID: "s1"},
+			want:    stuckSignalReclaimable,
+		},
+		{
+			name:    "office task is surface-only, never reclaimable",
+			task:    &models.Task{ID: "t1", WorkflowStepID: stepID, IsFromOffice: true},
+			session: &models.TaskSession{ID: "s1"},
+			want:    stuckSignalSurfaceOnly,
+		},
+		{
+			name:    "passthrough session is excluded even when the task is office",
+			task:    &models.Task{ID: "t1", WorkflowStepID: stepID, IsFromOffice: true},
+			session: &models.TaskSession{ID: "s1", IsPassthrough: true},
+			want:    stuckSignalNotCandidate,
+		},
+		{
+			name:    "passthrough session is excluded on an ordinary task",
+			task:    &models.Task{ID: "t1", WorkflowStepID: stepID},
+			session: &models.TaskSession{ID: "s1", IsPassthrough: true},
+			want:    stuckSignalNotCandidate,
+		},
+		{
+			name:    "stale signal on an office task is not reported as a stall",
+			task:    &models.Task{ID: "t1", WorkflowStepID: "step2", IsFromOffice: true},
+			session: &models.TaskSession{ID: "s1"},
+			want:    stuckSignalNotCandidate,
+		},
+		{
+			name:    "stale signal on an ordinary task is left to the signal-consuming paths",
+			task:    &models.Task{ID: "t1", WorkflowStepID: "step2"},
+			session: &models.TaskSession{ID: "s1"},
+			want:    stuckSignalNotCandidate,
+		},
+		{
+			name:    "nil task",
+			task:    nil,
+			session: &models.TaskSession{ID: "s1"},
+			want:    stuckSignalNotCandidate,
+		},
+		{
+			name:    "nil session",
+			task:    &models.Task{ID: "t1", WorkflowStepID: stepID},
+			session: nil,
+			want:    stuckSignalNotCandidate,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := stuckSignalDispositionFor(tc.task, tc.session, signal); got != tc.want {
+				t.Fatalf("disposition = %v, want %v", got, tc.want)
+			}
+		})
+	}
+}
+
+// An Office task must never be classified reclaimable, whatever else is true
+// of it. This is the invariant the whole unit rests on; if it ever fails, the
+// watchdog can forge a quorum decision.
+func TestStuckSignalDispositionFor_OfficeIsNeverReclaimable(t *testing.T) {
+	const stepID = "step1"
+	signal := models.PendingStepCompletionSignal{
+		StepID:     stepID,
+		SignaledAt: time.Now().UTC().Add(-30 * time.Minute),
+	}
+	for _, passthrough := range []bool{false, true} {
+		task := &models.Task{ID: "t1", WorkflowStepID: stepID, IsFromOffice: true}
+		session := &models.TaskSession{ID: "s1", IsPassthrough: passthrough}
+		if got := stuckSignalDispositionFor(task, session, signal); got == stuckSignalReclaimable {
+			t.Fatalf("office task classified reclaimable (passthrough=%v)", passthrough)
+		}
+	}
+}
+
+// seedOfficeStrandedSignal creates an Office task holding a step-completion
+// signal older than the watchdog threshold, in the given session state.
+func seedOfficeStrandedSignal(
+	t *testing.T, svc *Service, repo *sqliterepo.Repository,
+	taskID, sessionID, stepID string, state models.TaskSessionState,
+) models.PendingStepCompletionSignal {
+	t.Helper()
+	ctx := context.Background()
+
+	// models.Task.IsFromOffice is a derived projection, not a stored column
+	// (isFromOfficeProjection in the task repository): a task is Office when it
+	// has a non-empty project_id, or its workflow is the workspace's canonical
+	// office workflow. Setting the struct field would be silently discarded, so
+	// take the project_id branch.
+	if _, err := repo.DB().ExecContext(ctx,
+		`UPDATE tasks SET project_id = 'proj1' WHERE id = ?`, taskID); err != nil {
+		t.Fatalf("mark task as office: %v", err)
+	}
+	task, err := repo.GetTask(ctx, taskID)
+	if err != nil {
+		t.Fatalf("get task: %v", err)
+	}
+	if !task.IsFromOffice {
+		t.Fatalf("precondition: task %s should read as an Office task", taskID)
+	}
+
+	signal := models.PendingStepCompletionSignal{
+		StepID:     stepID,
+		Source:     models.StepCompletionSourceAgent,
+		Summary:    "all done",
+		SignaledAt: time.Now().UTC().Add(-30 * time.Minute),
+	}
+	if err := repo.SetSessionMetadataKey(ctx, sessionID, models.SessionMetaKeyPendingStepCompletion, signal); err != nil {
+		t.Fatalf("seed pending signal: %v", err)
+	}
+	if state != models.TaskSessionStateRunning {
+		if err := repo.UpdateTaskSessionState(ctx, sessionID, state, ""); err != nil {
+			t.Fatalf("set session state %q: %v", state, err)
+		}
+	}
+	return signal
+}
+
+func officeStrandedCount(gate stuckSignalGate) int64 {
+	v := officeStallStrandedSignalTotal.Get(officeStallLabel("gate", string(gate)))
+	iv, ok := v.(*expvar.Int)
+	if !ok {
+		return 0
+	}
+	return iv.Value()
+}
+
+// AC-001.2: an Office task entering through the stuckSignalCandidate gate is
+// surfaced, and is not reclaimed.
+func TestReclaimStuckSignalSessionsOnce_SurfacesOfficeTaskAtCandidateGate(t *testing.T) {
+	ctx := context.Background()
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	svc := createTestService(repo, officeStallStepGetter(), newMockTaskRepo())
+	seedOfficeStrandedSignal(t, svc, repo, "t1", "s1", "step1", models.TaskSessionStateRunning)
+
+	before := officeStrandedCount(stuckSignalGateCandidate)
+	svc.reclaimStuckSignalSessionsOnce(ctx)
+
+	if got := officeStrandedCount(stuckSignalGateCandidate) - before; got != 1 {
+		t.Fatalf("stranded-signal counter delta = %d, want 1", got)
+	}
+	assertOfficeStallNotActedOn(t, repo, "t1", "s1", "step1")
+}
+
+// Signal age alone is not inactivity (stuck_signal_watchdog.go's
+// stuckSignalWatchdogThreshold comment). An Office agent still producing
+// events within the inactivity window — a long tool call, a provider
+// retry/backoff — must not be surfaced just because its signal cleared the
+// age gate. Only the candidate gate needs this: the waiting gate's session is
+// already WAITING_FOR_INPUT, not running.
+func TestReclaimStuckSignalSessionsOnce_ActiveOfficeAgentNotSurfacedAtCandidateGate(t *testing.T) {
+	ctx := context.Background()
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	agentMgr := &mockAgentManager{
+		currentPromptExecutionID: "exec-1",
+	}
+	agentMgr.currentPromptLastActivityAt = time.Now().UTC().Add(-time.Minute)
+	svc := createTestServiceWithAgent(repo, officeStallStepGetter(), newMockTaskRepo(), agentMgr)
+	seedOfficeStrandedSignal(t, svc, repo, "t1", "s1", "step1", models.TaskSessionStateRunning)
+
+	before := officeStrandedCount(stuckSignalGateCandidate)
+	svc.reclaimStuckSignalSessionsOnce(ctx)
+
+	if got := officeStrandedCount(stuckSignalGateCandidate) - before; got != 0 {
+		t.Fatalf("stranded-signal counter delta = %d, want 0 — the agent is still active", got)
+	}
+	assertOfficeStallNotActedOn(t, repo, "t1", "s1", "step1")
+}
+
+// AC-001.3: an Office task entering through the
+// reconcileWaitingStuckSignalSessionIfDue gate is surfaced from that gate.
+// The two gates are near-identical predicates, so each needs its own proof:
+// covering only one leaves the other as a silent second gate.
+func TestReclaimStuckSignalSessionsOnce_SurfacesOfficeTaskAtWaitingGate(t *testing.T) {
+	ctx := context.Background()
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	svc := createTestService(repo, officeStallStepGetter(), newMockTaskRepo())
+	seedOfficeStrandedSignal(t, svc, repo, "t1", "s1", "step1", models.TaskSessionStateWaitingForInput)
+
+	before := officeStrandedCount(stuckSignalGateWaiting)
+	svc.reclaimStuckSignalSessionsOnce(ctx)
+
+	if got := officeStrandedCount(stuckSignalGateWaiting) - before; got != 1 {
+		t.Fatalf("stranded-signal counter delta = %d, want 1", got)
+	}
+	assertOfficeStallNotActedOn(t, repo, "t1", "s1", "step1")
+}
+
+// The waiting gate receives a session snapshot from the active-session scan.
+// A replacement signal on the same step must not inherit the old signal's age
+// while the detector is between that scan and its report.
+func TestReconcileWaitingStuckSignalSessionIfDue_SkipsReplacedSignal(t *testing.T) {
+	ctx := context.Background()
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	svc := createTestService(repo, officeStallStepGetter(), newMockTaskRepo())
+	oldSignal := seedOfficeStrandedSignal(t, svc, repo, "t1", "s1", "step1", models.TaskSessionStateWaitingForInput)
+	newSignal := models.PendingStepCompletionSignal{
+		StepID:     oldSignal.StepID,
+		Source:     models.StepCompletionSourceAgent,
+		SignaledAt: time.Now().UTC(),
+	}
+	if err := repo.SetSessionMetadataKey(ctx, "s1", models.SessionMetaKeyPendingStepCompletion, newSignal); err != nil {
+		t.Fatalf("replace pending signal: %v", err)
+	}
+
+	snapshot := &models.TaskSession{
+		ID:     "s1",
+		TaskID: "t1",
+		State:  models.TaskSessionStateWaitingForInput,
+		Metadata: map[string]interface{}{
+			models.SessionMetaKeyPendingStepCompletion: oldSignal,
+		},
+	}
+	before := officeStrandedCount(stuckSignalGateWaiting)
+	svc.reconcileWaitingStuckSignalSessionIfDue(ctx, snapshot, time.Now().UTC())
+	if got := officeStrandedCount(stuckSignalGateWaiting) - before; got != 0 {
+		t.Fatalf("stranded-signal counter delta = %d, want 0", got)
+	}
+}
+
+// AC-001.5: a still-stranded signal is reported once, not on every tick.
+func TestReclaimStuckSignalSessionsOnce_SurfacesOfficeSignalOncePerSignal(t *testing.T) {
+	ctx := context.Background()
+	repo := setupTestRepo(t)
+	seedSession(t, repo, "t1", "s1", "step1")
+	svc := createTestService(repo, officeStallStepGetter(), newMockTaskRepo())
+	seedOfficeStrandedSignal(t, svc, repo, "t1", "s1", "step1", models.TaskSessionStateRunning)
+
+	before := officeStrandedCount(stuckSignalGateCandidate)
+	for range 3 {
+		svc.reclaimStuckSignalSessionsOnce(ctx)
+	}
+	if got := officeStrandedCount(stuckSignalGateCandidate) - before; got != 1 {
+		t.Fatalf("counter delta across three scans = %d, want 1", got)
+	}
+
+	// A genuinely new signal on the same session and step reports again.
+	newer := models.PendingStepCompletionSignal{
+		StepID:     "step1",
+		Source:     models.StepCompletionSourceAgent,
+		SignaledAt: time.Now().UTC().Add(-20 * time.Minute),
+	}
+	if err := repo.SetSessionMetadataKey(ctx, "s1", models.SessionMetaKeyPendingStepCompletion, newer); err != nil {
+		t.Fatalf("reseed signal: %v", err)
+	}
+	svc.reclaimStuckSignalSessionsOnce(ctx)
+	if got := officeStrandedCount(stuckSignalGateCandidate) - before; got != 2 {
+		t.Fatalf("counter delta after a new signal = %d, want 2", got)
+	}
+}
+
+// AC-003.1 through AC-003.4: surfacing is the entire action.
+func assertOfficeStallNotActedOn(t *testing.T, repo *sqliterepo.Repository, taskID, sessionID, stepID string) {
+	t.Helper()
+	ctx := context.Background()
+
+	task, err := repo.GetTask(ctx, taskID)
+	if err != nil {
+		t.Fatalf("get task: %v", err)
+	}
+	if task.WorkflowStepID != stepID {
+		t.Fatalf("workflow step moved to %q; the watchdog applied a transition an Office quorum never approved", task.WorkflowStepID)
+	}
+	session, err := repo.GetTaskSession(ctx, sessionID)
+	if err != nil {
+		t.Fatalf("get session: %v", err)
+	}
+	if _, stillPending := models.LoadPendingStepSignal(session.Metadata); !stillPending {
+		t.Error("pending signal was consumed; surfacing must leave it for a human")
+	}
+}

apps/backend/internal/orchestrator/service.go

@@ -662,6 +662,30 @@ type Service struct {
 	// queuedMoveLifecycleLocks serializes source-exit work per task. The
 	// completion marker remains durable so a restart can safely resume work.
 	queuedMoveLifecycleLocks sync.Map
+	// officeStalledSignals dedupes the Office stranded-signal surfacing so a
+	// still-stranded signal is reported once, not on every 30-second reaper
+	// tick. Keyed by session ID, step ID and the signal's SignaledAt, so a
+	// genuinely new signal on the same session reports again. Deliberately
+	// in-memory: the surfaced state is derived and cheap to recompute, and a
+	// backend restart re-reporting each stranded signal once is preferable to
+	// a schema migration. Pruned against the live session set at the end of
+	// every complete scan, so it cannot outgrow the active-session list.
+	officeStalledSignals sync.Map
+	// officeDecisionWaiting dedupes the Office decision-waiting surfacing on
+	// the same terms as officeStalledSignals above. Keyed by task ID and step
+	// ID, so a task that moves on and later returns to a decision step (a
+	// rejection round) is reported again while a task sitting still is
+	// reported once. Pruned against the live candidate set at the end of
+	// every complete scan.
+	officeDecisionWaiting sync.Map
+	// officeStallDependenciesMu protects detector dependencies that are wired
+	// after Service.Start while the idle reaper can already read them.
+	officeStallDependenciesMu sync.RWMutex
+	// officeRunInFlight answers whether a task still has a queued or claimed
+	// run, the decision-waiting detector's false-positive guard. Nil until
+	// wired (SetOfficeRunInFlightReader); the detector reports nothing and
+	// counts a skip while it is.
+	officeRunInFlight officeRunInFlightReader
 	// engineOptions are applied each time initWorkflowEngine runs. Wired
 	// from cmd/kandev (Phase 3.2) to plug Phase 2 ADR-0004 dependencies
 	// — RunQueueAdapter, ParticipantStore, DecisionStore, and the CEO /
@@ -1104,6 +1128,16 @@ type Service struct {
 	ciAutomationWorkers sync.WaitGroup
 }
 
+func (s *Service) officeStallDependencies() (
+	engine.ParticipantStore,
+	engine.DecisionStore,
+	officeRunInFlightReader,
+) {
+	s.officeStallDependenciesMu.RLock()
+	defer s.officeStallDependenciesMu.RUnlock()
+	return s.engineParticipants, s.engineDecisions, s.officeRunInFlight
+}
+
 type RouteAction string
 
 const (
@@ -1816,14 +1850,18 @@ func (s *Service) SetEngineRunQueue(adapter engine.RunQueueAdapter) {
 
 // SetEngineParticipantStore wires the engine's ParticipantStore.
 func (s *Service) SetEngineParticipantStore(store engine.ParticipantStore) {
+	s.officeStallDependenciesMu.Lock()
 	s.engineParticipants = store
+	s.officeStallDependenciesMu.Unlock()
 	s.engineOptions = append(s.engineOptions, engine.WithParticipantStore(store))
 	s.reinitWorkflowEngine()
 }
 
 // SetEngineDecisionStore wires the engine's DecisionStore.
 func (s *Service) SetEngineDecisionStore(store engine.DecisionStore) {
+	s.officeStallDependenciesMu.Lock()
 	s.engineDecisions = store
+	s.officeStallDependenciesMu.Unlock()
 	s.engineOptions = append(s.engineOptions, engine.WithDecisionStore(store))
 	s.reinitWorkflowEngine()
 }

apps/backend/internal/orchestrator/stuck_signal_watchdog.go

@@ -44,7 +44,9 @@ import (
 // (handleRecoverableFailureLocked, PR #2963): only a session holding a
 // signal for the task's current step is touched. A long turn that never
 // signalled at all is out of scope and keeps running. Office sessions are
-// excluded — this path must not advance an Office task's step. Passthrough
+// excluded from the reclaim — this path must not advance an Office task's
+// step — but are still reported; see stuckSignalDispositionFor and
+// surfaceOfficeStalledSignal. Passthrough
 // (PTY) sessions are excluded too, mirroring the guard
 // flipStaleRunningToWaiting and markIdleAfterReset already apply
 // (event_handlers_workflow.go): a passthrough session manages its own
@@ -108,18 +110,28 @@ func (s *Service) reclaimStuckSignalSessionsOnce(ctx context.Context) {
 	scanCtx, cancel := context.WithTimeout(ctx, stuckSignalScanBudget)
 	defer cancel()
 	now := time.Now().UTC()
+	live := make(map[string]struct{}, len(sessions))
 	for i, session := range sessions {
 		if scanCtx.Err() != nil {
 			s.logger.Warn("stuck-signal watchdog: per-tick scan budget exceeded; deferring remaining candidates to the next tick",
 				zap.Int("candidates_scanned", i),
 				zap.Int("candidates_deferred", len(sessions)-i))
+			// Deliberately no prune: `live` is incomplete for a scan cut
+			// short here, and pruning against it would re-report signals on
+			// the sessions this tick never reached.
 			return
 		}
+		if session != nil {
+			if signal, ok := models.LoadPendingStepSignal(session.Metadata); ok {
+				live[officeStalledSignalKey(session.ID, signal)] = struct{}{}
+			}
+		}
 		if s.reconcileWaitingStuckSignalSessionIfDue(scanCtx, session, now) {
 			continue
 		}
 		s.reclaimStuckSignalSessionIfDue(scanCtx, session, now)
 	}
+	s.pruneOfficeStalledSignals(live)
 }
 
 // reconcileWaitingStuckSignalSessionIfDue retries an accepted signal after a
@@ -135,7 +147,29 @@ func (s *Service) reconcileWaitingStuckSignalSessionIfDue(ctx context.Context, s
 		return false
 	}
 	task, err := s.repo.GetTask(ctx, session.TaskID)
-	if err != nil || task == nil || task.WorkflowStepID != signal.StepID || task.IsFromOffice || session.IsPassthrough {
+	if err != nil || task == nil {
+		return false
+	}
+	switch stuckSignalDispositionFor(task, session, signal) {
+	case stuckSignalSurfaceOnly:
+		guard := s.lockCancelInFlightGuard(session.ID)
+		defer guard.release()
+		if s.isCancelInFlight(session.ID) {
+			return true
+		}
+		latest, latestSignal, stillPending := s.stuckSignalWaitingStillPending(ctx, session.ID, signal)
+		if !stillPending {
+			return true
+		}
+		latestTask, err := s.repo.GetTask(ctx, latest.TaskID)
+		if err != nil || latestTask == nil ||
+			stuckSignalDispositionFor(latestTask, latest, latestSignal) != stuckSignalSurfaceOnly {
+			return true
+		}
+		s.surfaceOfficeStalledSignal(latestTask, latest, latestSignal, stuckSignalGateWaiting, now)
+		return true
+	case stuckSignalReclaimable:
+	default:
 		return false
 	}
 	guard := s.lockCancelInFlightGuard(session.ID)
@@ -143,7 +177,7 @@ func (s *Service) reconcileWaitingStuckSignalSessionIfDue(ctx context.Context, s
 	if s.isCancelInFlight(session.ID) {
 		return true
 	}
-	_, stillPending := s.stuckSignalWaitingStillPending(ctx, session.ID, signal.StepID)
+	_, _, stillPending := s.stuckSignalWaitingStillPending(ctx, session.ID, signal)
 	if !stillPending {
 		return true
 	}
@@ -151,16 +185,20 @@ func (s *Service) reconcileWaitingStuckSignalSessionIfDue(ctx context.Context, s
 	return true
 }
 
-func (s *Service) stuckSignalWaitingStillPending(ctx context.Context, sessionID, stepID string) (*models.TaskSession, bool) {
+func (s *Service) stuckSignalWaitingStillPending(
+	ctx context.Context,
+	sessionID string,
+	expected models.PendingStepCompletionSignal,
+) (*models.TaskSession, models.PendingStepCompletionSignal, bool) {
 	latest, err := s.repo.GetTaskSession(ctx, sessionID)
 	if err != nil || latest == nil || latest.State != models.TaskSessionStateWaitingForInput {
-		return nil, false
+		return nil, models.PendingStepCompletionSignal{}, false
 	}
 	signal, ok := models.LoadPendingStepSignal(latest.Metadata)
-	if !ok || signal.StepID != stepID {
+	if !ok || officeStalledSignalKey(sessionID, signal) != officeStalledSignalKey(sessionID, expected) {
-		return nil, false
+		return nil, models.PendingStepCompletionSignal{}, false
 	}
-	return latest, true
+	return latest, signal, true
 }
 
 // reclaimStuckSignalSessionIfDue filters a single session against the
@@ -357,13 +395,10 @@ func (s *Service) prepareStuckSignalCancellation(
 }
 
 // stuckSignalCandidate applies the unguarded eligibility filter: RUNNING or
-// STARTING state, a pending signal older than stuckSignalWatchdogThreshold,
+// STARTING state, plus a pending signal older than
-// matching the task's CURRENT step, on a non-Office, non-passthrough task.
+// stuckSignalWatchdogThreshold that stuckSignalDispositionFor classifies as
-// Office sessions go FAILED, not WAITING_FOR_INPUT, and must not advance
+// reclaimable. See that function for why Office and passthrough sessions never
-// their step from this path — same exclusion PR #2963 applies to
+// reach the reclaim path.
-// handleRecoverableFailureLocked's reconciliation hook. Passthrough sessions
-// are excluded because this watchdog's inactivity gate cannot see them (see
-// the package doc and stuckSignalInactiveLongEnough).
 func (s *Service) stuckSignalCandidate(
 	ctx context.Context, session *models.TaskSession, now time.Time,
 ) (*models.Task, models.PendingStepCompletionSignal, bool) {
@@ -381,14 +416,145 @@ func (s *Service) stuckSignalCandidate(
 	if err != nil || task == nil {
 		return nil, models.PendingStepCompletionSignal{}, false
 	}
-	// Only a signal that matches the task's CURRENT step is this watchdog's
+	switch stuckSignalDispositionFor(task, session, signal) {
-	// job. A stale signal (step already moved on) is left for the ordinary
+	case stuckSignalSurfaceOnly:
-	// signal-consuming paths, which clear a bag entry that no longer
+		// Signal age alone is not inactivity (see stuckSignalWatchdogThreshold's
-	// matches the current step.
+		// comment above): an Office agent can still be mid-tool-call when its
-	if task.WorkflowStepID != signal.StepID || task.IsFromOffice || session.IsPassthrough {
+		// signal clears this gate. The waiting gate needs no equivalent check —
+		// that session is WAITING_FOR_INPUT, not running — but this gate's
+		// session is RUNNING/STARTING, so it is read here exactly as the
+		// reclaim path below would read it, just without registering a
+		// reclaim.
+		if _, inactive := s.stuckSignalInactiveLongEnough(ctx, session.ID, now); inactive {
+			s.surfaceOfficeStalledSignal(task, session, signal, stuckSignalGateCandidate, now)
+		}
+		return nil, models.PendingStepCompletionSignal{}, false
+	case stuckSignalReclaimable:
+		return task, signal, true
+	default:
 		return nil, models.PendingStepCompletionSignal{}, false
 	}
-	return task, signal, true
+}
+
+// stuckSignalDisposition is the verdict both watchdog gate sites consult. It
+// exists so the Office boundary is expressed exactly once: the two gates
+// (reconcileWaitingStuckSignalSessionIfDue and stuckSignalCandidate) carried
+// near-identical inline predicates, and a change applied to one of them
+// silently left the other as a second, divergent gate.
+type stuckSignalDisposition int
+
+const (
+	// stuckSignalNotCandidate: this watchdog has no business with the
+	// session at all.
+	stuckSignalNotCandidate stuckSignalDisposition = iota
+	// stuckSignalReclaimable: the session may be reclaimed - cancel the
+	// execution, force-close the turn, and apply the signalled transition.
+	stuckSignalReclaimable
+	// stuckSignalSurfaceOnly: the stuck state is real and worth reporting,
+	// but this watchdog must not act on it. Office tasks only.
+	stuckSignalSurfaceOnly
+)
+
+// stuckSignalDispositionFor classifies a session holding a pending signal.
+//
+// Office tasks are stuckSignalSurfaceOnly, never stuckSignalReclaimable. Both
+// gate sites lead to recovery: passing the
+// reconcileWaitingStuckSignalSessionIfDue gate falls through to
+// reconcileStepCompletionSignalLocked, which applies the step transition the
+// agent asked for, and passing the stuckSignalCandidate gate leads to
+// reclaimStuckSignalSessionOwned, which cancels the execution and force-closes
+// the turn before doing the same. Advancing an Office step from here would
+// record the transition without the decisions the step's quorum gate requires -
+// the forged approval those gates exist to prevent. Office therefore stays
+// excluded from recovery, and is lifted only for detection.
+//
+// Passthrough (PTY) sessions are stuckSignalNotCandidate regardless of Office
+// status: none of the ACP-only lastActivityAt writers ever run for one, so
+// there is no activity signal to distinguish a stuck passthrough session from a
+// busy one. Excluding them from detection as well as recovery keeps this
+// watchdog from reporting a state it cannot actually observe.
+func stuckSignalDispositionFor(
+	task *models.Task, session *models.TaskSession, signal models.PendingStepCompletionSignal,
+) stuckSignalDisposition {
+	if task == nil || session == nil {
+		return stuckSignalNotCandidate
+	}
+	// A stale signal (step already moved on) is left for the ordinary
+	// signal-consuming paths, which clear a bag entry that no longer matches
+	// the current step. Checked before the Office branch so a stale Office
+	// signal is not reported as a stall.
+	if task.WorkflowStepID != signal.StepID {
+		return stuckSignalNotCandidate
+	}
+	if session.IsPassthrough {
+		return stuckSignalNotCandidate
+	}
+	if task.IsFromOffice {
+		return stuckSignalSurfaceOnly
+	}
+	return stuckSignalReclaimable
+}
+
+// stuckSignalGate names which of the two watchdog gate sites observed a
+// stranded signal, so the report says where it was seen.
+type stuckSignalGate string
+
+const (
+	stuckSignalGateWaiting   stuckSignalGate = "waiting_reconcile"
+	stuckSignalGateCandidate stuckSignalGate = "candidate"
+)
+
+// officeStalledSignalKey identifies one stranded signal for dedupe purposes.
+// SignaledAt is part of the key so a genuinely new signal on the same session
+// and step is reported again rather than suppressed by its predecessor.
+func officeStalledSignalKey(sessionID string, signal models.PendingStepCompletionSignal) string {
+	return sessionID + "\x00" + signal.StepID + "\x00" + signal.SignaledAt.UTC().Format(time.RFC3339Nano)
+}
+
+// surfaceOfficeStalledSignal reports an Office task holding a stranded
+// completion signal, and does nothing else. No session write, no turn close,
+// no step transition, no decision row, no queued run — surfacing this state is
+// the whole action, because repairing it would apply a step transition the
+// task's quorum gate never approved.
+func (s *Service) surfaceOfficeStalledSignal(
+	task *models.Task,
+	session *models.TaskSession,
+	signal models.PendingStepCompletionSignal,
+	gate stuckSignalGate,
+	now time.Time,
+) {
+	if task == nil || session == nil {
+		return
+	}
+	if _, seen := s.officeStalledSignals.LoadOrStore(officeStalledSignalKey(session.ID, signal), struct{}{}); seen {
+		return
+	}
+	officeStallStrandedSignalTotal.Add(officeStallLabel("gate", string(gate)), 1)
+	s.logger.Warn("office stall watchdog: Office task holds a stranded completion signal",
+		zap.String("task_id", task.ID),
+		zap.String("session_id", session.ID),
+		zap.String("step_id", signal.StepID),
+		zap.Duration("signal_age", now.Sub(signal.SignaledAt)),
+		zap.String("gate", string(gate)))
+}
+
+// pruneOfficeStalledSignals drops dedupe entries for signals that are no
+// longer present on any active session, bounding the map by the active-session
+// list. Only ever called after a scan that visited every session: a scan cut
+// short by the budget has an incomplete live set, and pruning against it would
+// re-report signals that are still stranded.
+func (s *Service) pruneOfficeStalledSignals(live map[string]struct{}) {
+	s.officeStalledSignals.Range(func(key, _ any) bool {
+		k, ok := key.(string)
+		if !ok {
+			s.officeStalledSignals.Delete(key)
+			return true
+		}
+		if _, stillLive := live[k]; !stillLive {
+			s.officeStalledSignals.Delete(k)
+		}
+		return true
+	})
 }
 
 // stuckSignalStillPending re-reads the session under the cancelInFlight

apps/backend/internal/task/models/models.go

@@ -920,6 +920,27 @@ func (t *Task) IsOfficeOwnedAndAssigned() bool {
 	return t != nil && t.IsFromOffice && t.AssigneeAgentProfileID != ""
 }
 
+// OfficeDecisionWaitCandidate is the compact projection the Office
+// decision-waiting detector scans (REQ-OFFICE-STALL-VISIBILITY-002). It is a
+// candidate, not a finding: the repository query only establishes that the
+// task is Office-owned, sits at a step carrying a decision-required seat, and
+// has been quiet since UpdatedAt. Whether a decision was already recorded and
+// whether a run is still in flight are judged by the detector, so each
+// rejection has its own countable reason.
+type OfficeDecisionWaitCandidate struct {
+	TaskID    string    `db:"id"`
+	StepID    string    `db:"workflow_step_id"`
+	UpdatedAt time.Time `db:"updated_at"`
+}
+
+// OfficeDecisionWaitCursor identifies the last candidate in one ordered page.
+// The task repository uses it to continue a bounded scan without repeatedly
+// returning the same oldest rows.
+type OfficeDecisionWaitCursor struct {
+	UpdatedAt time.Time
+	TaskID    string
+}
+
 // ChildCompletionRow is the compact active-child projection used to decide
 // whether a parent task's on_children_completed trigger is ready to fire.
 type ChildCompletionRow struct {

apps/backend/internal/task/repository/sqlite/base_migrations.go

@@ -107,6 +107,7 @@ func (r *Repository) runMigrations() error {
 		return err
 	}
 	r.migrate.Apply("idx_tasks_queued_for_step", `CREATE INDEX IF NOT EXISTS idx_tasks_queued_for_step ON tasks(queued_for_step_id, queued_at)`)
+	r.migrate.Apply("idx_tasks_updated_at_id", `CREATE INDEX IF NOT EXISTS idx_tasks_updated_at_id ON tasks(updated_at, id)`)
 	// Remove deprecated workflow_step_id column from task_sessions
 	if err := r.migrateSessionsRemoveWorkflowStepID(); err != nil {
 		return err

apps/backend/internal/task/repository/sqlite/base_schema.go

@@ -553,6 +553,7 @@ func (r *Repository) initCoreIndexes() error {
 	CREATE INDEX IF NOT EXISTS idx_tasks_workflow_id ON tasks(workflow_id);
 	CREATE INDEX IF NOT EXISTS idx_tasks_workflow_step_id ON tasks(workflow_step_id);
 	CREATE INDEX IF NOT EXISTS idx_tasks_archived_at ON tasks(archived_at);
+	CREATE INDEX IF NOT EXISTS idx_tasks_updated_at_id ON tasks(updated_at, id);
 	CREATE INDEX IF NOT EXISTS idx_task_repositories_task_id ON task_repositories(task_id);
 	CREATE INDEX IF NOT EXISTS idx_task_repositories_repository_id ON task_repositories(repository_id);
 	CREATE INDEX IF NOT EXISTS idx_task_workspace_folders_task_position ON task_workspace_folders(task_id, position);

apps/backend/internal/task/repository/sqlite/office_decision_candidates.go

@@ -0,0 +1,97 @@
+package sqlite
+
+import (
+	"context"
+	"fmt"
+	"time"
+
+	"github.com/kandev/kandev/internal/agentctl/tracing"
+	"github.com/kandev/kandev/internal/task/models"
+)
+
+// officeDecisionWaitScanLimit bounds one page. The detector runs on the idle
+// reaper's tick and does per-candidate follow-up reads, so an unbounded result
+// set would let a large parked backlog stretch a single tick past the scan
+// budget. Keyset pagination lets one tick continue with later rows until that
+// budget expires.
+const officeDecisionWaitScanLimit = 200
+
+// ListOfficeDecisionWaitCandidates returns Office tasks that sit at a workflow
+// step carrying a decision-required seat and have not been touched since
+// quietSince. Pass after to continue after a previous ordered page. The return
+// cursor is nil when the result set is complete.
+//
+// This is deliberately only the cheap, indexable half of the
+// decision-waiting predicate (REQ-OFFICE-STALL-VISIBILITY-002). It does not
+// look at workflow_step_decisions or at the runs queue: those are judged by
+// the orchestrator's detector so a candidate rejected for "a decision already
+// exists" is distinguishable from one rejected for "a run is still in flight",
+// and so an unreadable input can fail closed with its own counted reason
+// instead of silently shrinking a SQL result set.
+//
+// Office identity comes from IsFromOfficePredicate — the same expression that
+// backs models.Task.IsFromOffice — rather than a hand-rolled project_id test,
+// because IsFromOffice is a read-time projection and a second, drifting
+// definition of "Office task" is exactly the bug that would reintroduce the
+// exclusion this feature exists to remove.
+//
+// tasks.updated_at is the quiet-time anchor. It is an approximation of "when
+// the task entered this step": an unrelated edit to the task row (a title
+// change, a reorder) restarts the clock and delays surfacing. That error is
+// one-directional — it under-surfaces, never over-surfaces — which is the
+// right direction for a detector whose stated non-goal is acting on what it
+// finds. The exact anchor is task_step_transitions.occurred_at, which has no
+// read API yet; tightening to it is a follow-up, not a correction.
+func (r *Repository) ListOfficeDecisionWaitCandidates(
+	ctx context.Context,
+	quietSince time.Time,
+	after *models.OfficeDecisionWaitCursor,
+) ([]models.OfficeDecisionWaitCandidate, *models.OfficeDecisionWaitCursor, error) {
+	ctx, span := tracing.Tracer("kandev-db").Start(ctx, "db.ListOfficeDecisionWaitCandidates")
+	defer span.End()
+
+	cursorPredicate := ""
+	args := []interface{}{quietSince}
+	if after != nil {
+		cursorPredicate = `
+		  AND (t.updated_at > ? OR (t.updated_at = ? AND t.id > ?))`
+		args = append(args, after.UpdatedAt, after.UpdatedAt, after.TaskID)
+	}
+	query := fmt.Sprintf(`
+		SELECT t.id, t.workflow_step_id, t.updated_at
+		FROM tasks t
+		WHERE t.archived_at IS NULL
+		  AND COALESCE(t.workflow_step_id, '') != ''
+		  AND t.state NOT IN ('COMPLETED', 'FAILED', 'CANCELLED')
+		  AND t.updated_at < ?
+		%s
+		  AND %s
+		  AND %s
+		  AND EXISTS (
+			SELECT 1 FROM workflow_step_participants p
+			WHERE p.step_id = t.workflow_step_id
+			  AND p.decision_required = 1
+			  AND (COALESCE(p.task_id, '') = '' OR p.task_id = t.id)
+		  )
+		ORDER BY t.updated_at ASC, t.id ASC
+		LIMIT %d
+	`,
+		cursorPredicate,
+		isFromOfficeProjection("t"),
+		excludeConfigModePredicate(r.ro.DriverName(), "t.metadata"),
+		officeDecisionWaitScanLimit,
+	)
+
+	var out []models.OfficeDecisionWaitCandidate
+	if err := r.ro.SelectContext(ctx, &out, r.ro.Rebind(query), args...); err != nil {
+		return nil, nil, fmt.Errorf("list office decision-wait candidates: %w", err)
+	}
+	if len(out) == officeDecisionWaitScanLimit {
+		last := out[len(out)-1]
+		return out, &models.OfficeDecisionWaitCursor{
+			UpdatedAt: last.UpdatedAt,
+			TaskID:    last.TaskID,
+		}, nil
+	}
+	return out, nil, nil
+}

apps/backend/internal/task/repository/sqlite/office_decision_candidates_postgres_test.go

@@ -0,0 +1,60 @@
+package sqlite
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/kandev/kandev/internal/task/models"
+	"github.com/kandev/kandev/internal/testutil"
+)
+
+// TestPostgresListOfficeDecisionWaitCandidates is the PostgreSQL twin of the
+// SQLite candidate-enumeration tests. The query is dialect-sensitive in two
+// places the SQLite run cannot exercise: excludeConfigModePredicate expands to
+// a JSON expression that differs per driver, and the office-identity EXISTS
+// subquery joins workspaces on a nullable column. A regression in either is a
+// query that errors on every tick in production while every SQLite test stays
+// green — the detector would then surface nothing and look healthy.
+// Skips unless KANDEV_TEST_POSTGRES_DSN is set.
+func TestPostgresListOfficeDecisionWaitCandidates(t *testing.T) {
+	db := testutil.OpenIsolatedPostgres(t, testutil.PostgresDSNFromEnv(t))
+	repo, err := NewWithDB(db, db, nil)
+	if err != nil {
+		t.Fatalf("init postgres schema: %v", err)
+	}
+	ctx := context.Background()
+
+	const workspaceID = "ws-pg-decision-wait"
+	if err := repo.CreateWorkspace(ctx, &models.Workspace{
+		ID: workspaceID, Name: "pg-decision-wait", OwnerID: "u-1",
+	}); err != nil {
+		t.Fatalf("CreateWorkspace: %v", err)
+	}
+	officeWorkflowID, err := repo.EnsureOfficeWorkflow(ctx, workspaceID)
+	if err != nil {
+		t.Fatalf("EnsureOfficeWorkflow: %v", err)
+	}
+
+	seatRow(t, repo, "pg-seat-review", "pg-step-review", "", "reviewer", true)
+	seatRow(t, repo, "pg-seat-plain", "pg-step-plain", "", "reviewer", false)
+
+	seedCandidateTask(t, repo, workspaceID, officeWorkflowID, "pg-t-waiting", "pg-step-review", "", "IN_PROGRESS")
+	seedCandidateTask(t, repo, workspaceID, "wf-pg-kanban", "pg-t-kanban", "pg-step-review", "", "IN_PROGRESS")
+	seedCandidateTask(t, repo, workspaceID, officeWorkflowID, "pg-t-no-seat", "pg-step-plain", "", "IN_PROGRESS")
+
+	quiet := time.Now().UTC().Add(-2 * time.Hour)
+	for _, id := range []string{"pg-t-waiting", "pg-t-kanban", "pg-t-no-seat"} {
+		stampTaskQuiet(t, repo, id, quiet)
+	}
+
+	got := candidateIDs(t, repo, time.Now().UTC().Add(-time.Hour))
+	if !hasID(got, "pg-t-waiting") {
+		t.Errorf("pg-t-waiting missing from candidates %v", got)
+	}
+	for _, unwanted := range []string{"pg-t-kanban", "pg-t-no-seat"} {
+		if hasID(got, unwanted) {
+			t.Errorf("candidate %q must not be selected; got %v", unwanted, got)
+		}
+	}
+}

apps/backend/internal/task/repository/sqlite/office_decision_candidates_test.go

@@ -0,0 +1,209 @@
+package sqlite
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/kandev/kandev/internal/task/models"
+	v1 "github.com/kandev/kandev/pkg/api/v1"
+)
+
+// seatRow inserts a workflow_step_participants row. taskID "" is a
+// template-level seat (applies to every task at the step); a non-empty taskID
+// is a per-task override.
+func seatRow(t *testing.T, repo *Repository, id, stepID, taskID, role string, decisionRequired bool) {
+	t.Helper()
+	required := 0
+	if decisionRequired {
+		required = 1
+	}
+	if _, err := repo.db.ExecContext(context.Background(), repo.db.Rebind(`
+		INSERT INTO workflow_step_participants
+			(id, step_id, task_id, role, agent_profile_id, decision_required, position)
+		VALUES (?, ?, ?, ?, 'agent-1', ?, 0)
+	`), id, stepID, taskID, role, required); err != nil {
+		t.Fatalf("seed seat %s: %v", id, err)
+	}
+}
+
+// stampTaskQuiet backdates a task's updated_at. CreateTask stamps "now", and
+// the candidate query's whole purpose is to select on that column, so every
+// fixture has to be aged explicitly.
+func stampTaskQuiet(t *testing.T, repo *Repository, taskID string, at time.Time) {
+	t.Helper()
+	if _, err := repo.db.ExecContext(context.Background(),
+		repo.db.Rebind(`UPDATE tasks SET updated_at = ? WHERE id = ?`), at, taskID); err != nil {
+		t.Fatalf("backdate %s: %v", taskID, err)
+	}
+}
+
+func candidateIDs(t *testing.T, repo *Repository, quietSince time.Time) []string {
+	t.Helper()
+	var (
+		rows   []models.OfficeDecisionWaitCandidate
+		cursor *models.OfficeDecisionWaitCursor
+	)
+	for {
+		page, next, err := repo.ListOfficeDecisionWaitCandidates(context.Background(), quietSince, cursor)
+		if err != nil {
+			t.Fatalf("ListOfficeDecisionWaitCandidates: %v", err)
+		}
+		rows = append(rows, page...)
+		if next == nil {
+			break
+		}
+		cursor = next
+	}
+	ids := make([]string, 0, len(rows))
+	for _, row := range rows {
+		ids = append(ids, row.TaskID)
+	}
+	return ids
+}
+
+func hasID(ids []string, want string) bool {
+	for _, id := range ids {
+		if id == want {
+			return true
+		}
+	}
+	return false
+}
+
+// officeCandidateFixture builds one workspace with a materialised office
+// workflow and returns (repo, workspaceID, officeWorkflowID).
+func officeCandidateFixture(t *testing.T) (*Repository, string, string) {
+	t.Helper()
+	repo := newRepoForBuiltinWorkflowTests(t)
+	ctx := context.Background()
+	const workspaceID = "ws-decision-wait"
+	if err := repo.CreateWorkspace(ctx, &models.Workspace{
+		ID: workspaceID, Name: "decision-wait", OwnerID: "u-1",
+	}); err != nil {
+		t.Fatalf("CreateWorkspace: %v", err)
+	}
+	officeWorkflowID, err := repo.EnsureOfficeWorkflow(ctx, workspaceID)
+	if err != nil {
+		t.Fatalf("EnsureOfficeWorkflow: %v", err)
+	}
+	return repo, workspaceID, officeWorkflowID
+}
+
+func seedCandidateTask(
+	t *testing.T, repo *Repository, workspaceID, workflowID, taskID, stepID, projectID string,
+	state string,
+) {
+	t.Helper()
+	task := &models.Task{
+		ID:             taskID,
+		WorkspaceID:    workspaceID,
+		WorkflowID:     workflowID,
+		WorkflowStepID: stepID,
+		Title:          taskID,
+		State:          v1.TaskState(state),
+		ProjectID:      projectID,
+	}
+	if err := repo.CreateTask(context.Background(), task); err != nil {
+		t.Fatalf("CreateTask %s: %v", taskID, err)
+	}
+}
+
+// TestListOfficeDecisionWaitCandidates_SelectsOnlyOfficeTasksAtDecisionSeats
+// pins the cheap half of REQ-OFFICE-STALL-VISIBILITY-002's predicate. Each
+// negative row differs from the positive one in exactly one respect, so a
+// broken clause fails a named case rather than merely shrinking a list.
+func TestListOfficeDecisionWaitCandidates_SelectsOnlyOfficeTasksAtDecisionSeats(t *testing.T) {
+	repo, workspaceID, officeWorkflowID := officeCandidateFixture(t)
+	quiet := time.Now().UTC().Add(-2 * time.Hour)
+
+	const (
+		reviewStep = "step-review"
+		plainStep  = "step-plain"
+	)
+	seatRow(t, repo, "seat-review", reviewStep, "", "reviewer", true)
+	seatRow(t, repo, "seat-plain", plainStep, "", "reviewer", false)
+
+	// Positive: office-by-workflow, decision-required seat, quiet.
+	seedCandidateTask(t, repo, workspaceID, officeWorkflowID, "t-waiting", reviewStep, "", "IN_PROGRESS")
+	// Office by project link rather than workflow — must also qualify.
+	seedCandidateTask(t, repo, workspaceID, "wf-kanban", "t-waiting-project", reviewStep, "proj-1", "IN_PROGRESS")
+	// Kanban task at the same step: excluded by the Office predicate.
+	seedCandidateTask(t, repo, workspaceID, "wf-kanban", "t-kanban", reviewStep, "", "IN_PROGRESS")
+	// Office task at a step whose only seat is not decision-required.
+	seedCandidateTask(t, repo, workspaceID, officeWorkflowID, "t-no-decision-seat", plainStep, "", "IN_PROGRESS")
+	// Office task at a step with no seats at all.
+	seedCandidateTask(t, repo, workspaceID, officeWorkflowID, "t-seatless", "step-none", "", "IN_PROGRESS")
+	// Office task in a terminal state.
+	seedCandidateTask(t, repo, workspaceID, officeWorkflowID, "t-done", reviewStep, "", "COMPLETED")
+
+	for _, id := range []string{
+		"t-waiting", "t-waiting-project", "t-kanban",
+		"t-no-decision-seat", "t-seatless", "t-done",
+	} {
+		stampTaskQuiet(t, repo, id, quiet)
+	}
+
+	got := candidateIDs(t, repo, time.Now().UTC().Add(-time.Hour))
+
+	for _, want := range []string{"t-waiting", "t-waiting-project"} {
+		if !hasID(got, want) {
+			t.Errorf("candidate %q missing from %v", want, got)
+		}
+	}
+	for _, unwanted := range []string{"t-kanban", "t-no-decision-seat", "t-seatless", "t-done"} {
+		if hasID(got, unwanted) {
+			t.Errorf("candidate %q must not be selected; got %v", unwanted, got)
+		}
+	}
+}
+
+func TestListOfficeDecisionWaitCandidates_RecentlyTouchedTaskIsNotQuiet(t *testing.T) {
+	repo, workspaceID, officeWorkflowID := officeCandidateFixture(t)
+	seatRow(t, repo, "seat-review", "step-review", "", "approver", true)
+	seedCandidateTask(t, repo, workspaceID, officeWorkflowID, "t-fresh", "step-review", "", "IN_PROGRESS")
+	stampTaskQuiet(t, repo, "t-fresh", time.Now().UTC().Add(-5*time.Minute))
+
+	got := candidateIDs(t, repo, time.Now().UTC().Add(-time.Hour))
+	if hasID(got, "t-fresh") {
+		t.Errorf("a task touched 5 minutes ago must not be a candidate at a 1h threshold; got %v", got)
+	}
+}
+
+func TestListOfficeDecisionWaitCandidates_PerTaskSeatQualifiesOnlyItsOwnTask(t *testing.T) {
+	repo, workspaceID, officeWorkflowID := officeCandidateFixture(t)
+	quiet := time.Now().UTC().Add(-2 * time.Hour)
+
+	// The only decision-required seat at this step is a per-task override
+	// belonging to t-mine, so t-theirs sits at a step with no seat of its own.
+	seatRow(t, repo, "seat-mine", "step-review", "t-mine", "reviewer", true)
+	seedCandidateTask(t, repo, workspaceID, officeWorkflowID, "t-mine", "step-review", "", "IN_PROGRESS")
+	seedCandidateTask(t, repo, workspaceID, officeWorkflowID, "t-theirs", "step-review", "", "IN_PROGRESS")
+	stampTaskQuiet(t, repo, "t-mine", quiet)
+	stampTaskQuiet(t, repo, "t-theirs", quiet)
+
+	got := candidateIDs(t, repo, time.Now().UTC().Add(-time.Hour))
+	if !hasID(got, "t-mine") {
+		t.Errorf("t-mine holds the per-task decision seat and must be a candidate; got %v", got)
+	}
+	if hasID(got, "t-theirs") {
+		t.Errorf("t-theirs has no seat of its own and must not borrow t-mine's; got %v", got)
+	}
+}
+
+func TestListOfficeDecisionWaitCandidates_ArchivedTaskIsExcluded(t *testing.T) {
+	repo, workspaceID, officeWorkflowID := officeCandidateFixture(t)
+	seatRow(t, repo, "seat-review", "step-review", "", "reviewer", true)
+	seedCandidateTask(t, repo, workspaceID, officeWorkflowID, "t-archived", "step-review", "", "IN_PROGRESS")
+	archivedAt := time.Now().UTC().Add(-3 * time.Hour)
+	if _, err := repo.db.ExecContext(context.Background(),
+		repo.db.Rebind(`UPDATE tasks SET archived_at = ?, updated_at = ? WHERE id = ?`),
+		archivedAt, archivedAt, "t-archived"); err != nil {
+		t.Fatalf("archive: %v", err)
+	}
+
+	got := candidateIDs(t, repo, time.Now().UTC().Add(-time.Hour))
+	if hasID(got, "t-archived") {
+		t.Errorf("an archived task must not be surfaced as decision-waiting; got %v", got)
+	}
+}

apps/backend/internal/workflow/adapters/adapters.go

@@ -164,6 +164,7 @@ func (a *DecisionAdapter) ListStepDecisions(
 			DeciderID:     d.DeciderID,
 			Role:          d.Role,
 			Comment:       d.Comment,
+			SupersededAt:  d.SupersededAt,
 		})
 	}
 	return out, nil

apps/backend/internal/workflow/engine/adapters.go

@@ -92,6 +92,13 @@ type DecisionInfo struct {
 	DeciderID   string
 	Role        string
 	Comment     string
+
+	// SupersededAt is non-nil once a rework round has superseded this row
+	// (workflow_step_decisions.superseded_at). ListStepDecisions returns
+	// superseded rows alongside active ones so timelines render full history;
+	// a caller that needs only the currently-decided state must filter on
+	// this field itself rather than treating a non-empty result as "decided".
+	SupersededAt *time.Time
 }
 
 // ParticipantStore reads the workflow_step_participants table for an engine

docs/plans/office-stall-visibility/plan.md

@@ -0,0 +1,135 @@
+---
+spec: docs/specs/office/requirements/stall-visibility.md
+design: docs/specs/office/system-design/stall-visibility.md
+created: 2026-09-01
+status: done
+---
+
+# Implementation Plan: Office Stall Visibility
+
+## Overview
+
+Make two invisible Office stall states visible from the backend, without giving
+any sweep permission to repair them. Work lands in three waves: restructure the
+watchdog so its Office boundary is expressed once (Wave 1), add stranded-signal
+surfacing for Office tasks (Wave 2), and add the decision-waiting detector
+(Wave 3).
+
+## Confirmed root cause
+
+Verified against the working tree on 2026-09-01.
+
+`internal/orchestrator/stuck_signal_watchdog.go` excludes Office at two sites:
+
+- **:138** — `reconcileWaitingStuckSignalSessionIfDue`
+- **:388** — `stuckSignalCandidate`
+
+Both are **recovery** gates, not detection gates. Passing either leads to a
+reclaim that cancels the agent execution, force-closes the turn, and calls
+`reconcileStepCompletionSignalLocked` to apply the requested step transition.
+The originating plan unit describes these exclusions as the defect and asks for
+them to be lifted, while separately forbidding any automatic transition for
+Office. Both cannot hold: lifting the term at either site is what causes the
+automatic transition. This plan resolves the conflict in favour of the stated
+non-goal — detection is lifted for Office, recovery is not.
+
+The second gap is real and unqualified. No detector covers a task parked at a
+review or approval step with no run in flight and no decision recorded. The
+Office inbox's `inboxTaskReviewRequests` surfaces "a decision is outstanding",
+but has no age threshold and no in-flight-run condition, so a healthy task being
+actively worked is indistinguishable from a stalled one.
+
+## Schema findings that shape the work
+
+Confirmed before implementation, as the plan unit asked:
+
+- The `runs` table has **no `task_id` column**. Task linkage is
+  `payload -> '$.task_id'`. Use `dialect.JSONExtract(driver, "payload",
+  "task_id")`, not a literal `json_extract` — Postgres is a supported driver.
+  Precedent: `internal/office/repository/sqlite/failure.go:351` with
+  `failure_postgres_test.go`.
+- Run statuses are `queued`, `claimed`, `finished`, `failed`, `cancelled`. There
+  is no `running`; **`claimed` is the executing state**. "No queued or running
+  run" maps to `status IN ('queued','claimed')`. Precedent for the whole
+  predicate: `ReapStaleCheckouts` in `internal/office/repository/sqlite/tasks.go`.
+- `orchestrator.Service` already holds step-scoped `engineParticipants`
+  (`ListStepParticipants`) and `engineDecisions` (`ListStepDecisions`). No new
+  wiring is needed for seats or decisions.
+- `engineRunQueue` is **write-only** (`QueueRun`). A new narrow reader interface
+  is required for the in-flight-run term.
+- The Office repository's `ListActiveTaskDecisions` is deliberately task-scoped
+  across all steps. Use the engine's step-scoped `ListStepDecisions` instead.
+
+---
+
+## Backend
+
+### Wave 1: express the Office boundary once
+
+Replace the two inline `task.IsFromOffice` terms with a single
+`stuckSignalDisposition` helper returning `notCandidate` / `reclaimable` /
+`surfaceOnly`. Pure refactor: Office still returns a non-recovery outcome at both
+sites, so behavior is unchanged and the existing suite must stay green.
+
+### Wave 2: surface a stranded signal on an Office task
+
+Add `surfaceOfficeStalledSignal`, called on the `surfaceOnly` outcome from both
+sites, with an in-memory dedupe keyed by session, step, and `SignaledAt`. Emits
+a `Warn` log carrying the observing gate, and increments
+`office_stall_stranded_signal_total`.
+
+### Wave 3: decision-waiting detector
+
+Add `office_decision_stall_watchdog.go` with a second pass on the existing reaper
+tick, plus `officeRunInFlightReader` and its Office repository implementation,
+wired from `internal/backendapp`. Predicate order puts the run check last, since it is the
+most expensive term. Every unreadable input fails closed to "do not surface".
+
+## Waves
+
+Wave 1:
+
+- [x] [Task 01: Express the watchdog's Office boundary once](task-01-single-office-disposition.md)
+
+Wave 2:
+
+- [x] [Task 02: Surface an Office task's stranded signal](task-02-surface-stranded-signal.md)
+
+Wave 3:
+
+- [x] [Task 03: Detect an Office task waiting on a decision](task-03-decision-waiting-detector.md)
+
+## Required workflow verification
+
+After all targeted task checks pass:
+
+1. Run `make -C apps/backend fmt`.
+2. Run `make -C apps/backend test lint`.
+3. Commit the explicit changed paths with a Conventional Commit message.
+
+## Risks and non-goals
+
+- **The reclaim path must stay closed to Office.** Wave 1 is the load-bearing
+  change: if a later edit routes `surfaceOnly` into the reclaim branch, an Office
+  step advances without its quorum. The Wave 1 test asserting that no Office task
+  reaches the reclaim path from either site is the guard against that, and must
+  not be weakened.
+- **A wrong alert is the failure mode that matters here.** Nothing is repaired,
+  so the only way this capability can do harm is by training operators to ignore
+  it. Every unreadable input fails closed, and
+  `office_stall_detector_skipped_total` exists so a silently degraded detector is
+  distinguishable from a quiet system.
+- **The decision-waiting threshold is provisional and deliberately separate.**
+  Wave 3 uses its own `officeDecisionWaitingThreshold`, set to 60 minutes, and
+  must not reuse `stuckSignalWatchdogThreshold`. Ten minutes of session silence
+  is anomalous; ten minutes of a human not yet deciding is a reviewer getting
+  coffee. Sharing the constant would also couple two unrelated phenomena, so
+  tuning session silence would silently retune this. Sixty minutes is expected to
+  move once there is field data.
+- **No user-visible surface in this plan.** Detection lands as logs and expvar
+  counters only. Promoting it to an Office inbox row is deliberately deferred: a
+  new inbox type carries a frontend change plus copy in five languages, and it
+  should be designed against real detector output rather than in advance.
+- No schema migration, no new table, no new HTTP endpoint, no change to
+  non-Office reclaim behavior, and no retirement of the workspace-scoped
+  `Stall Session Watchdog` automation.

docs/plans/office-stall-visibility/task-01-single-office-disposition.md

@@ -0,0 +1,54 @@
+---
+id: "01-single-office-disposition"
+title: "Express the watchdog's Office boundary once"
+status: done
+wave: 1
+depends_on: []
+plan: "plan.md"
+spec: "../../specs/office/requirements/stall-visibility.md"
+---
+
+# Task 01: Express the watchdog's Office boundary once
+
+Pure refactor. No behavior change; the existing suite must stay green.
+
+## Acceptance
+
+- A single `stuckSignalDisposition` helper is the only place `task.IsFromOffice`
+  is evaluated in `stuck_signal_watchdog.go`, returning `stuckSignalNotCandidate`,
+  `stuckSignalReclaimable`, or `stuckSignalSurfaceOnly`.
+- `reconcileWaitingStuckSignalSessionIfDue` (currently :138) and
+  `stuckSignalCandidate` (currently :388) both consult it; neither retains an
+  inline `IsFromOffice` term.
+- An Office task returns `stuckSignalSurfaceOnly` from both sites and reaches
+  neither reclaim path.
+- A non-Office task's reclaim behavior is byte-for-byte unchanged.
+- Passthrough sessions still return `stuckSignalNotCandidate` from both sites.
+
+## Verification
+
+- `make -C apps/backend test` — the whole existing
+  `internal/orchestrator` suite, including
+  `stuck_signal_watchdog_test.go` and
+  `stuck_signal_watchdog_cancellation_test.go`, passes unchanged.
+- New table-driven test asserting the disposition for each of: Office task,
+  non-Office task, passthrough session, step mismatch, signal too young.
+- New test asserting an Office task reaches neither reclaim path, entered
+  independently through each of the two sites. This is the guard named in the
+  plan's risk section and must not be weakened.
+
+Write the disposition tests first against the current inline predicates so they
+pass before and after the refactor; that is what proves it is behavior-preserving.
+
+## Files likely touched
+
+- `apps/backend/internal/orchestrator/stuck_signal_watchdog.go`
+- `apps/backend/internal/orchestrator/stuck_signal_watchdog_test.go`
+
+## Dependencies
+
+None.
+
+## Parallelism
+
+Blocks Tasks 02 and 03. Both build on the disposition helper.

docs/plans/office-stall-visibility/task-02-surface-stranded-signal.md

@@ -0,0 +1,58 @@
+---
+id: "02-surface-stranded-signal"
+title: "Surface an Office task's stranded signal"
+status: done
+wave: 2
+depends_on: ["01-single-office-disposition"]
+plan: "plan.md"
+spec: "../../specs/office/requirements/stall-visibility.md"
+---
+
+# Task 02: Surface an Office task's stranded signal
+
+Satisfies REQ-OFFICE-STALL-VISIBILITY-001.
+
+## Acceptance
+
+- `surfaceOfficeStalledSignal` fires on the `stuckSignalSurfaceOnly` outcome from
+  both gate sites, emitting one `Warn` log with `task_id`, `session_id`,
+  `step_id`, `signal_age`, and `gate`.
+- `office_stall_stranded_signal_total` increments, labelled by `gate`, using the
+  `expvar.NewMap` + `metricLabel` convention in
+  `internal/office/scheduler/metrics_vars.go`.
+- An in-memory dedupe keyed by session ID, step ID, and the signal's `SignaledAt`
+  suppresses re-surfacing across the 30-second reaper ticks; the entry is dropped
+  when the signal clears.
+- Nothing else changes: no session state write, no turn close, no step
+  transition, no decision row, no queued run.
+
+## Verification
+
+- `make -C apps/backend test`
+- Test: an Office task with a stranded signal surfaces exactly once across three
+  consecutive scans.
+- Test: entered through `stuckSignalCandidate`, it surfaces (AC-001.2).
+- Test: entered through `reconcileWaitingStuckSignalSessionIfDue`, it surfaces
+  (AC-001.3).
+- Test: after surfacing, task step, session state, decision rows and run queue
+  are all unchanged (AC-003.1 through AC-003.4).
+- Test: a new signal on the same session with a later `SignaledAt` surfaces
+  again.
+
+Each test must first fail on the Task 01 tree, where the surface-only outcome is
+silent.
+
+## Files likely touched
+
+- `apps/backend/internal/orchestrator/stuck_signal_watchdog.go`
+- `apps/backend/internal/orchestrator/stuck_signal_watchdog_test.go`
+- `apps/backend/internal/orchestrator/service.go` (dedupe map field)
+- new: `apps/backend/internal/orchestrator/office_stall_metrics.go`
+
+## Dependencies
+
+Task 01.
+
+## Parallelism
+
+Independent of Task 03 once Task 01 lands.

docs/plans/office-stall-visibility/task-03-decision-waiting-detector.md

@@ -0,0 +1,71 @@
+---
+id: "03-decision-waiting-detector"
+title: "Detect an Office task waiting on a decision"
+status: done
+wave: 3
+depends_on: ["01-single-office-disposition"]
+plan: "plan.md"
+spec: "../../specs/office/requirements/stall-visibility.md"
+---
+
+# Task 03: Detect an Office task waiting on a decision
+
+Satisfies REQ-OFFICE-STALL-VISIBILITY-002.
+
+## Acceptance
+
+- A second detection pass runs on the existing idle-reaper tick, bounded by the
+  same `stuckSignalScanBudget` and sharing its defer-to-next-tick behavior.
+- Predicate, evaluated in this order: task is Office → current step has a
+  `decision_required` reviewer or approver seat
+  (`engineParticipants.ListStepParticipants`) → no active decision for
+  (task, current step) (`engineDecisions.ListStepDecisions`) → age over threshold
+  → no in-flight run. The run check is last because it is the most expensive.
+- `officeRunInFlightReader.HasInFlightRunForTask` is implemented in the Office
+  SQLite repository using `dialect.JSONExtract(driver, "payload", "task_id")` and
+  `status IN ('queued','claimed')`, and is wired from `internal/backendapp`, where the Office repository is constructed.
+- Surfacing is a `Warn` log plus `office_stall_decision_waiting_total`, with the
+  same in-memory dedupe (keyed by task ID and step ID).
+- Every unreadable input fails closed to "do not surface" and increments
+  `office_stall_detector_skipped_total` labelled by `reason`.
+- Threshold is its own named constant, `officeDecisionWaitingThreshold` = 60
+  minutes, commented as provisional. Reusing `stuckSignalWatchdogThreshold` is
+  explicitly rejected: it couples two unrelated phenomena.
+
+## Verification
+
+- `make -C apps/backend test`
+- Test: a decision-waiting Office task past the threshold is surfaced
+  (AC-002.1).
+- Test: **the same task with a `claimed` run is not surfaced** (AC-002.2). This
+  is the false-positive guard named in the definition of done.
+- Test: the same task with a `queued` run is not surfaced.
+- Test: an active decision row for the current step suppresses surfacing
+  (AC-002.3).
+- Test: a step with only non-decision-required or watcher/collaborator seats is
+  not surfaced (AC-002.4).
+- Test: run reader unwired, and run reader returning an error, each skip and
+  count (AC-002.5).
+- Test: a non-Office task is never evaluated (AC-002.6).
+- Test: surfacing writes no decision row, queues no run, and does not change the
+  task's step (AC-003.1 through AC-003.3).
+- Postgres-dialect test for the new repository query, mirroring
+  `internal/office/repository/sqlite/failure_postgres_test.go`.
+
+## Files likely touched
+
+- new: `apps/backend/internal/orchestrator/office_decision_stall_watchdog.go`
+- new: `apps/backend/internal/orchestrator/office_decision_stall_watchdog_test.go`
+- `apps/backend/internal/orchestrator/service.go` (reader setter)
+- `apps/backend/internal/orchestrator/idle_session_reaper.go` (tick call site)
+- `apps/backend/internal/office/repository/sqlite/runs.go`
+- `apps/backend/internal/office/repository/sqlite/runs_test.go`
+- `apps/backend/internal/backendapp/main.go` (wiring)
+
+## Dependencies
+
+Task 01. Independent of Task 02.
+
+## Parallelism
+
+Can run alongside Task 02.

docs/specs/office/README.md

@@ -52,6 +52,7 @@ dashboard projections, and Office testing contracts.
 - [Office Review Participant Seats](requirements/review-participant-seats.md)
 - [Office Agent Runtime — Error Handling Contract](requirements/runtime.md)
 - [Office Scheduler](requirements/scheduler.md)
+- [Office Stall Visibility](requirements/stall-visibility.md)
 - [Step Entry Sequence Execution](requirements/step-entry-sequence-execution.md)
 - [Office Tasks](requirements/tasks.md)
 - [Office: E2E Mock Harness for Task Sessions and Messages](requirements/testing.md)
@@ -99,6 +100,7 @@ dashboard projections, and Office testing contracts.
 - [Office Agent Runtime — Error Handling Contract System Design Part 1](system-design/runtime-01.md)
 - [Office Agent Runtime — Error Handling Contract System Design Part 2](system-design/runtime-02.md)
 - [Office Scheduler System Design Part 1](system-design/scheduler-01.md)
+- [Office Stall Visibility System Design](system-design/stall-visibility.md)
 - [Step Entry Sequence Execution System Design](system-design/step-entry-sequence-execution.md)
 - [Office Scheduler System Design Part 2](system-design/scheduler-02.md)
 - [Office Tasks System Design Part 1](system-design/tasks-01.md)

docs/specs/office/requirements/stall-visibility.md

@@ -0,0 +1,140 @@
+---
+status: active
+system: office
+created: 2026-09-01
+owners:
+  - kandev
+---
+
+# Office Stall Visibility Requirements
+
+## Overview
+
+An Office task can stop making progress in two ways that no current backend
+detector reports. First, an Office session can accept a step-completion signal
+and never reach turn-end; the stuck-signal watchdog
+(`internal/orchestrator/stuck_signal_watchdog.go`) excludes every Office task by
+construction, so the state is invisible. Second, an Office task can sit at a
+review or approval step with a decision-required seat, no decision recorded, and
+no run in flight; nothing is stalled in the session sense, because there is no
+session, so no session-oriented detector can see it.
+
+The Office system owns this contract because the stuck state is defined by
+Office primitives: `tasks.is_from_office`, the participant slate, the decision
+ledger, and the Office run queue. The orchestrator's stuck-signal watchdog is an
+adjacent contract this capability reads and constrains, but does not own.
+
+This capability makes both states **visible**. It does not recover either one.
+
+## Terminology
+
+- **Stranded signal:** a pending `step_complete_kandev` signal recorded in a
+  session's metadata bag that matches the task's current step and has not been
+  applied.
+- **Reclaim:** the stuck-signal watchdog's recovery action, which cancels the
+  agent execution, force-closes the stale turn, moves the session to
+  `WAITING_FOR_INPUT`, and applies the requested step transition through
+  `reconcileStepCompletionSignalLocked`.
+- **Surface:** report a stuck state to an operator or user without changing task,
+  session, decision, or run state.
+- **Decision-waiting task:** an Office task whose current step carries at least
+  one `decision_required` reviewer or approver seat with no active decision row
+  for that task and step.
+- **In-flight run:** a row in the Office `runs` table for the task with status
+  `queued` or `claimed`. The table has no terminal `running` status; `claimed`
+  is the executing state.
+
+## Requirements
+
+### REQ-OFFICE-STALL-VISIBILITY-001: Surface an Office task holding a stranded signal
+
+**Intent:** An Office task whose session accepted a completion signal and then
+went quiet is currently invisible to every backend detector. Make it visible
+without granting the watchdog permission to advance an Office step.
+
+**User story:** As an operator, I want a stranded completion signal on an Office
+task to be reported, so that I can intervene before the card sits indefinitely.
+
+The stuck-signal watchdog applies its Office exclusion at two independent sites
+(`reconcileWaitingStuckSignalSessionIfDue` and `stuckSignalCandidate`). Both are
+**recovery** gates: passing either one leads to a reclaim. Removing the
+`IsFromOffice` term from either predicate would make the watchdog cancel an
+Office agent and apply the step transition it requested, which is the forged
+quorum decision that REQ-OFFICE-STALL-VISIBILITY-003 forbids. The exclusion is
+therefore preserved for recovery and lifted only for detection.
+
+#### Acceptance criteria
+
+- **AC-OFFICE-STALL-VISIBILITY-001.1:** When an Office task's session holds a
+  stranded signal older than the watchdog threshold, the system shall surface it
+  exactly once per distinct signal, and shall not reclaim the session.
+- **AC-OFFICE-STALL-VISIBILITY-001.2:** When an Office task reaches the
+  `stuckSignalCandidate` evaluation path, the system shall surface the stranded
+  signal from that path.
+- **AC-OFFICE-STALL-VISIBILITY-001.3:** When an Office task reaches the
+  `reconcileWaitingStuckSignalSessionIfDue` evaluation path, the system shall
+  surface the stranded signal from that path.
+- **AC-OFFICE-STALL-VISIBILITY-001.4:** When a task is not an Office task, the
+  system shall preserve the existing reclaim behavior with no change in outcome.
+- **AC-OFFICE-STALL-VISIBILITY-001.5:** While a surfaced signal remains pending
+  and unchanged, the system shall not re-surface it on subsequent scans.
+
+### REQ-OFFICE-STALL-VISIBILITY-002: Surface a decision-waiting Office task
+
+**Intent:** The most common Office stall is a task parked at a review or approval
+step with nobody coming. No session is stuck, so no session-oriented detector
+observes it.
+
+**User story:** As an operator, I want an Office task that has been waiting on a
+decision with no work in flight to be reported, so that a human can decide it.
+
+#### Acceptance criteria
+
+- **AC-OFFICE-STALL-VISIBILITY-002.1:** When an Office task's current step
+  carries at least one `decision_required` reviewer or approver seat, has no
+  active decision row for that task and step, has no in-flight run, and has been
+  in that state for longer than the configured threshold, the system shall
+  surface the task as decision-waiting.
+- **AC-OFFICE-STALL-VISIBILITY-002.2:** When the task has an in-flight run, the
+  system shall not surface it, regardless of decision state or age.
+- **AC-OFFICE-STALL-VISIBILITY-002.3:** When an active decision row exists for
+  the task and its current step, the system shall not surface it.
+- **AC-OFFICE-STALL-VISIBILITY-002.4:** When the current step carries no
+  `decision_required` reviewer or approver seat, the system shall not surface it.
+- **AC-OFFICE-STALL-VISIBILITY-002.5:** When the in-flight run state cannot be
+  read, the system shall not surface the task and shall record the skip.
+- **AC-OFFICE-STALL-VISIBILITY-002.6:** When a task is not an Office task, the
+  system shall not evaluate it for this detector.
+
+### REQ-OFFICE-STALL-VISIBILITY-003: Never act on a surfaced Office stall
+
+**Intent:** Office quorum gates exist to stop a step advancing without the
+decisions it requires. A detector that repaired what it found would forge those
+decisions.
+
+**User story:** As a workspace owner, I want stall detection to report and never
+remediate, so that an automated sweep cannot manufacture an approval.
+
+#### Acceptance criteria
+
+- **AC-OFFICE-STALL-VISIBILITY-003.1:** When either detector surfaces an Office
+  task, the system shall not change the task's workflow step.
+- **AC-OFFICE-STALL-VISIBILITY-003.2:** When either detector surfaces an Office
+  task, the system shall not write a decision row.
+- **AC-OFFICE-STALL-VISIBILITY-003.3:** When either detector surfaces an Office
+  task, the system shall not queue a run, cancel an agent, or close a turn.
+- **AC-OFFICE-STALL-VISIBILITY-003.4:** When either detector surfaces an Office
+  task, the system shall not change any session state.
+
+## Out of scope
+
+- Automatic recovery, step advancement, decision synthesis, or run queueing for
+  any Office stall. This is the explicit non-goal above.
+- Changing reclaim behavior for non-Office tasks. The existing stuck-signal
+  watchdog contract is unchanged for them.
+- Making `lastActivityAt` meaningful for passthrough sessions. Passthrough
+  remains excluded from both detection and recovery.
+- Detecting a long-running Office turn that never signalled at all. That is the
+  in-process stall ticker's domain.
+- Replacing the workspace-scoped `Stall Session Watchdog` automation. Retiring it
+  is a follow-up once backend coverage is confirmed in the field.

docs/specs/office/system-design/stall-visibility.md

@@ -0,0 +1,257 @@
+---
+status: current
+system: office
+requirements:
+  - REQ-OFFICE-STALL-VISIBILITY-001
+  - REQ-OFFICE-STALL-VISIBILITY-002
+  - REQ-OFFICE-STALL-VISIBILITY-003
+---
+
+# Office Stall Visibility System Design
+
+## Purpose and boundaries
+
+This design adds two detection-only passes to the orchestrator's existing
+stuck-signal watchdog scan. The Office system owns the outcome because the
+states are defined by Office primitives; the orchestrator owns the scan loop and
+the session lifecycle, which this design reads and constrains but does not
+change for non-Office tasks.
+
+Adjacent contracts used and not owned:
+
+- `internal/orchestrator/stuck_signal_watchdog.go` — the scan loop, the
+  threshold, the cancellation guards, and the reclaim path.
+- `internal/workflow/engine` `ParticipantStore` / `DecisionStore` — already wired
+  onto `orchestrator.Service` as `engineParticipants` / `engineDecisions`, both
+  step-scoped and nil-safe.
+- `internal/office/repository/sqlite` — the `runs` queue.
+- `internal/task/repository/sqlite` — the `tasks` table, the
+  `workflow_step_participants` table, and `IsFromOfficePredicate`, the exported
+  SQL expression that defines Office ownership.
+
+The detector lives beside the existing watchdog rather than in the Office
+scheduler for two reasons: the step-scoped participant and decision readers this
+detector needs are already wired onto `orchestrator.Service` (the Office
+repository's `ListActiveTaskDecisions` is deliberately task-scoped across all
+steps), and co-locating keeps one place to look for stall detection.
+
+## Requirement mapping
+
+| Requirement | Design section |
+| --- | --- |
+| `REQ-OFFICE-STALL-VISIBILITY-001` | [Splitting detection from recovery](#splitting-detection-from-recovery) |
+| `REQ-OFFICE-STALL-VISIBILITY-002` | [Decision-waiting detector](#decision-waiting-detector) |
+| `REQ-OFFICE-STALL-VISIBILITY-003` | [Failure and recovery](#failure-and-recovery) |
+
+## Components and responsibilities
+
+- **`stuckSignalDisposition`** (new, `stuck_signal_watchdog.go`) — the single
+  predicate both existing gate sites consult. Returns one of three outcomes and
+  is the only place the `IsFromOffice` term is evaluated.
+- **`surfaceOfficeStalledSignal`** (new) — emits the stranded-signal surfacing
+  for an Office task and records the dedupe key.
+- **`detectOfficeDecisionWaitingOnce`** (new file,
+  `office_decision_stall_watchdog.go`) — the second detection pass, driven from
+  the same reaper tick.
+- **`officeDecisionCandidateLister`** (new narrow optional interface) — the
+  candidate enumeration, implemented by the task repository as
+  `ListOfficeDecisionWaitCandidates`.
+- **`officeRunInFlightReader`** (new narrow optional interface) — reports whether
+  a task has a `queued` or `claimed` run.
+- **Office SQLite repository** — implements the reader with a dialect-portable
+  JSON predicate.
+
+## Data and contracts
+
+### Shared disposition
+
+```go
+type stuckSignalDisposition int
+
+const (
+    stuckSignalNotCandidate stuckSignalDisposition = iota
+    stuckSignalReclaimable
+    stuckSignalSurfaceOnly // Office: visible, never reclaimed
+)
+```
+
+Both `reconcileWaitingStuckSignalSessionIfDue` and
+`stuckSignalCandidate` drop their inline
+`task.IsFromOffice` term and call this helper instead. Routing both sites through
+one predicate is what makes AC-001.2 and AC-001.3 structurally true rather than
+separately maintained: the two predicates are near-identical today, and fixing
+one in isolation would leave the other as a silent second gate.
+
+`stuckSignalSurfaceOnly` is returned for an Office task that satisfies every
+other candidacy term. Callers treat it as "not a recovery candidate" — the same
+control flow the current `false` return produces — after calling
+`surfaceOfficeStalledSignal`.
+
+Passthrough sessions continue to return `stuckSignalNotCandidate` from both
+paths. The inactivity gate cannot read them, so they are excluded from detection
+as well as recovery.
+
+### Run-queue reader
+
+```go
+type officeRunInFlightReader interface {
+    HasInFlightRunForTask(ctx context.Context, taskID string) (bool, error)
+}
+```
+
+Follows the optional-interface idiom already used by `stuckSignalSessionLister`
+and `promptActivitySessionReader`: asserted at use, nil-safe, so lightweight test
+doubles stay compatible. Wired from `internal/backendapp`, where the Office
+repository is constructed, keeping the orchestrator decoupled from the Office
+packages.
+
+The repository implementation must use
+`dialect.JSONExtract(r.ro.DriverName(), "payload", "task_id")` rather than a
+literal `json_extract`. The `runs` table has no `task_id` column — task linkage
+lives in the payload JSON — and Postgres is a supported driver. Precedent:
+`internal/office/repository/sqlite/failure.go` and its
+`failure_postgres_test.go`. The status predicate is `status IN ('queued',
+'claimed')`; the table has no `running` status.
+
+### Decision and participant reads
+
+- Seats: `engineParticipants.ListStepParticipants(ctx, stepID, taskID)`, filtered
+  to `DecisionRequired` seats with role `reviewer` or `approver`.
+- Decisions: `engineDecisions.ListStepDecisions(ctx, taskID, stepID)`.
+
+Both are step-scoped, which is what AC-002.1 requires. `ListStepDecisions`
+returns superseded rows alongside active ones by its own documented contract
+(quorum guards filter at the call site), and `SupersedeTaskDecisions` leaves a
+reworked task's prior decision rows in place with `superseded_at` set rather
+than deleting or re-scoping them. The detector must therefore filter on
+`DecisionInfo.SupersededAt` itself — treating any non-empty result as
+"decided" reads a re-entered step as permanently decided after the first
+rework round, which is exactly the repeat stall this detector exists to find.
+
+## Control flow
+
+Both passes run inside `reclaimStuckSignalSessionsOnce`, on the idle reaper's
+existing 30-second ticker. No new goroutine.
+
+1. Existing per-session loop. For each session, the disposition helper decides
+   `notCandidate` / `reclaimable` / `surfaceOnly`. `surfaceOnly` surfaces and
+   continues; `reclaimable` follows the unchanged reclaim path.
+2. A second pass enumerates candidates, then applies the remaining
+   decision-waiting predicate to each: current step has a decision-required
+   reviewer/approver seat → no decision recorded for (task, step) → no in-flight
+   run. Order matters: the run check is the most expensive term and is evaluated
+   last.
+
+The two passes are siblings on the reaper tick, not nested: the decision-waiting
+pass is a separate call in `startIdleSessionReaper`'s tick closure, because a
+task waiting on a decision need not have any session at all and so cannot be
+found by iterating sessions.
+
+The decision-waiting pass is bounded by the same `stuckSignalScanBudget` already
+guarding the tick, and shares its "defer the remainder to the next tick"
+behavior so it cannot starve `reclaimIdleSessionsOnce`.
+
+### Candidate enumeration
+
+The driving query lives in the task repository, because it reads `tasks` and
+must define Office ownership with `IsFromOfficePredicate` — the same expression
+that backs `models.Task.IsFromOffice`. A second, hand-rolled definition of
+"Office task" would drift, which is exactly the class of bug this feature
+exists to remove.
+
+```go
+ListOfficeDecisionWaitCandidates(ctx, quietSince) ([]models.OfficeDecisionWaitCandidate, error)
+```
+
+It selects only the cheap, indexable half of the predicate: not archived, not in
+a terminal state, at a step carrying a `decision_required` seat (template-level
+or that task's own override), Office-owned, not a config-mode task, and
+`updated_at < quietSince`. Everything else is judged in the orchestrator so each
+rejection has its own countable reason and each unreadable input can fail closed
+independently. The result is ordered oldest-first and capped, so a truncated
+scan drops the freshest candidates and the next tick still sees them.
+
+**Age anchor.** `tasks.updated_at` is an approximation of "when the task entered
+this step". An unrelated edit to the task row restarts the clock and delays
+surfacing. That error is one-directional — it under-surfaces, never
+over-surfaces — which is the correct direction for a detector whose stated
+non-goal is acting on what it finds. The exact anchor is
+`task_step_transitions.occurred_at`, which has no read API yet; tightening to it
+is a follow-up, not a correction.
+
+### Threshold
+
+The decision-waiting pass uses its own constant,
+`officeDecisionWaitingThreshold`, set provisionally to 60 minutes. It must not
+reuse `stuckSignalWatchdogThreshold`. The two measure unrelated phenomena: a
+session that has emitted no events for ten minutes is anomalous, whereas a
+decision waiting on a human for ten minutes is ordinary. Sharing the constant
+would also mean a future tuning of session silence silently retunes
+decision-waiting. Sixty minutes is still plausibly normal but worth counting; a
+first alert that cries wolf gets muted, and a muted detector is worth nothing.
+
+## Failure and recovery
+
+This capability has no recovery path by design (REQ-OFFICE-STALL-VISIBILITY-003).
+Every failure mode resolves to "do not surface, record the skip":
+
+- Reader unwired or type assertion fails → skip. Absence of the run reader must
+  never be read as absence of runs; that would surface every decision-waiting
+  task including healthy ones.
+- Reader returns an error → skip and count. Unknown is not "no run in flight".
+- Participant or decision store unwired, or either read failing → skip and
+  count. Mirrors the engine's existing `ReasonParticipantStoreUnwired` /
+  `ReasonDecisionStoreUnwired` discipline.
+- Candidate lister unwired, or the enumeration query failing → skip the whole
+  tick and count. The next tick retries.
+
+The asymmetry with the existing watchdog is deliberate. The reclaim path fails
+closed because a wrong reclaim corrupts a live session; these passes fail closed
+because a wrong alert trains operators to ignore the signal.
+
+## Persistence
+
+No schema change. No migration. No new table.
+
+The stranded-signal dedupe key is in-memory on `orchestrator.Service`, keyed by
+session ID, step ID, and the signal's `SignaledAt`, and dropped when the signal
+clears or the session is reclaimed. A backend restart re-surfaces each still-
+stranded signal once; that is acceptable and preferable to a schema change,
+because the surfaced state is itself derived and cheap to recompute.
+
+The decision-waiting pass is stateless per tick. Re-surfacing is suppressed by
+the same in-memory dedupe keyed by task ID and step ID.
+
+## Security
+
+No new trust boundary. Both passes read state the orchestrator already reads
+within the same process, and neither exposes a new endpoint. Surfaced payloads
+carry task, session, and step identifiers plus a duration; they must not carry
+prompt text, agent output, or decision notes.
+
+## Observability
+
+Structured logs, at `Warn`, on the existing watchdog logger:
+
+- `office stall watchdog: Office task holds a stranded completion signal` with
+  `task_id`, `session_id`, `step_id`, `signal_age`, and `gate` (which of the two
+  sites observed it).
+- `office stall watchdog: Office task is waiting on a decision that has not
+  been recorded` with `task_id`, `step_id`, `task_updated_at`, `quiet_for`, and
+  `threshold`.
+
+Counters, following the `expvar.NewMap` + `metricLabel` convention already used
+by `internal/office/scheduler/metrics_vars.go` and exposed through `/debug/vars`:
+
+- `office_stall_stranded_signal_total`, labelled by `gate`.
+- `office_stall_decision_waiting_total`.
+- `office_stall_detector_skipped_total`, labelled by `reason`
+  (`run_reader_unwired`, `run_reader_error`, `participant_store_unwired`,
+  `decision_store_unwired`, `participant_read_failed`, `decision_read_failed`,
+  `candidate_lister_unwired`, `candidate_list_failed`), so a silently degraded
+  detector is distinguishable from a genuinely quiet system.
+
+## Related decisions
+
+None. This design adds no architectural decision beyond the existing watchdog
+contract; the Office recovery exclusion it preserves was established by PR #2963.