Show agent liveness, stuck/zombie detection and parse errors in admin

internal/agents/manager.go

@@ -50,6 +50,9 @@ type Manager struct {
 
 	deliverMu sync.Mutex
 	capMu     sync.Mutex
+
+	bad       *badLines
+	watchStop chan struct{}
 }
 
 type live struct {
@@ -82,6 +85,8 @@ func New(st *store.Store, limits Limits, worktreeRoot string) *Manager {
 		live:         map[string]*live{},
 		rateLimits:   map[api.AgentSessionVendor]map[string]any{},
 		merges:       map[string]string{},
+		bad:          newBadLines(),
+		watchStop:    make(chan struct{}),
 	}
 }
 
@@ -262,6 +267,8 @@ func (m *Manager) Recover(ctx context.Context) error {
 }
 
 func (m *Manager) Close() {
+	close(m.watchStop)
+
 	m.mu.Lock()
 	m.closing = true
 	all := []*live{}
@@ -559,6 +566,7 @@ func (m *Manager) ensureLive(ctx context.Context, sess api.AgentSession, task ap
 			"UserPromptSubmit": {{Func: remindGuidance}},
 			"Stop":             {{Func: m.bounceProseQuestion(l)}},
 		},
+		OnBadLine: func(line []byte, reason string) { m.recordBadLine(l, line, reason) },
 	}
 	started, err := m.store.SessionStarted(ctx, sess.Id)
 	if err != nil {
@@ -595,6 +603,9 @@ func (m *Manager) pump(l *live) {
 		if err != nil {
 			slog.Error("record agent event", "session", l.id, "type", ev.Type, "err", err)
 		}
+		if isDecodeError(err) {
+			m.recordBadLine(l, ev.Raw, fmt.Sprintf("decode %s: %s", ev.Type, err))
+		}
 		if took := time.Since(start); took > 200*time.Millisecond {
 			slog.Info("agent event slow", "session", l.id, "type", ev.Type, "took", took)
 		}