refactor(monitor): direct event push, no filesystem polling

CompletionNotifier interface now has AgentStarted() and
AgentCompleted() instead of Poke(). Dispatch pushes notifications
directly to monitor with agent/repo/status data. Monitor pushes
MCP channel events immediately — no scanning, no dedup maps,
no filesystem polling latency.

Events.jsonl kept as audit log only, not notification mechanism.
Timer-based scan kept for startup seeding and stale detection.

Co-Authored-By: Virgil <virgil@lethean.io>

pkg/agentic/dispatch.go

@@ -145,8 +145,16 @@ func (s *PrepSubsystem) spawnAgent(agent, prompt, wsDir string) (int, string, er
 	proc.CloseStdin()
 	pid := proc.Info().PID
 
-	// Emit start event for channel notifications
-	emitStartEvent(agent, core.PathBase(wsDir))
+	// Notify monitor directly — no filesystem polling
+	if s.onComplete != nil {
+		st, _ := readStatus(wsDir)
+		repo := ""
+		if st != nil {
+			repo = st.Repo
+		}
+		s.onComplete.AgentStarted(agent, repo, core.PathBase(wsDir))
+	}
+	emitStartEvent(agent, core.PathBase(wsDir)) // audit log
 
 	go func() {
 		ticker := time.NewTicker(5 * time.Second)
@@ -190,10 +198,16 @@ func (s *PrepSubsystem) spawnAgent(agent, prompt, wsDir string) (int, string, er
 			writeStatus(wsDir, st)
 		}
 
-		emitCompletionEvent(agent, core.PathBase(wsDir), finalStatus)
+		emitCompletionEvent(agent, core.PathBase(wsDir), finalStatus) // audit log
 
+		// Push notification directly — no filesystem polling
 		if s.onComplete != nil {
-			s.onComplete.Poke()
+			stNow, _ := readStatus(wsDir)
+			repoName := ""
+			if stNow != nil {
+				repoName = stNow.Repo
+			}
+			s.onComplete.AgentCompleted(agent, repoName, core.PathBase(wsDir), finalStatus)
 		}
 
 		if finalStatus == "completed" {

pkg/agentic/prep.go

@@ -22,12 +22,13 @@ import (
 	"gopkg.in/yaml.v3"
 )
 
-// CompletionNotifier is called when an agent completes, to trigger
-// immediate notifications to connected clients.
+// CompletionNotifier receives agent lifecycle events directly from dispatch.
+// No filesystem polling — events flow in-memory.
 //
 //	prep.SetCompletionNotifier(monitor)
 type CompletionNotifier interface {
-	Poke()
+	AgentStarted(agent, repo, workspace string)
+	AgentCompleted(agent, repo, workspace, status string)
 }
 
 // PrepSubsystem provides agentic MCP tools for workspace orchestration.

pkg/agentic/prep_test.go

@@ -194,9 +194,14 @@ func TestSetCompletionNotifier_Good(t *testing.T) {
 }
 
 type mockNotifier struct {
-	poked bool
+	started   bool
+	completed bool
 }
 
-func (m *mockNotifier) Poke() {
-	m.poked = true
+func (m *mockNotifier) AgentStarted(agent, repo, workspace string) {
+	m.started = true
+}
+
+func (m *mockNotifier) AgentCompleted(agent, repo, workspace, status string) {
+	m.completed = true
 }

pkg/monitor/monitor.go

@@ -214,9 +214,7 @@ func (m *Subsystem) Shutdown(_ context.Context) error {
 	return nil
 }
 
-// Poke triggers an immediate check cycle.
-//
-//	mon.Poke()
+// Poke triggers an immediate check cycle (legacy — prefer AgentStarted/AgentCompleted).
 func (m *Subsystem) Poke() {
 	select {
 	case m.poke <- struct{}{}:
@@ -224,6 +222,43 @@ func (m *Subsystem) Poke() {
 	}
 }
 
+// AgentStarted pushes an immediate notification when an agent spawns.
+// Called directly by dispatch — no filesystem polling.
+//
+//	mon.AgentStarted("codex:gpt-5.3-codex-spark", "go-io", "core/go-io/task-5")
+func (m *Subsystem) AgentStarted(agent, repo, workspace string) {
+	if m.notifier != nil {
+		m.notifier.ChannelSend(context.Background(), "agent.started", map[string]any{
+			"agent": agent,
+			"repo":  repo,
+		})
+	}
+}
+
+// AgentCompleted pushes an immediate notification when an agent finishes.
+// Called directly by dispatch — no filesystem polling needed.
+//
+//	mon.AgentCompleted("codex", "go-io", "core/go-io/task-5", "completed")
+func (m *Subsystem) AgentCompleted(agent, repo, workspace, status string) {
+	if m.notifier != nil {
+		// Count current running/queued from status for context
+		running := 0
+		queued := 0
+		m.mu.Lock()
+		m.seenCompleted[workspace] = true
+		m.mu.Unlock()
+
+		m.notifier.ChannelSend(context.Background(), "agent.complete", map[string]any{
+			"completed": []string{core.Sprintf("%s (%s) [%s]", repo, agent, status)},
+			"count":     1,
+			"running":   running,
+			"queued":    queued,
+		})
+	}
+	// Also poke to update counts for any other monitors
+	m.Poke()
+}
+
 func (m *Subsystem) loop(ctx context.Context) {
 	// Initial check after short delay (let server fully start)
 	select {