fix: trigger LLM execution after /plan start approval

/plan start was a fast-path command that changed MEMORY.md status to
"executing" but never enqueued a message for the LLM worker, so the
agent never actually started working on the plan.

Add planStartPending flag that causes Run() to enqueue a synthetic
message after the fast-path response, triggering an LLM iteration.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

pkg/agent/loop.go

@@ -38,9 +38,10 @@ type AgentLoop struct {
 	stats          *stats.Tracker // nil when --stats not passed
 	running        atomic.Bool
 	summarizing    sync.Map
-	fallback       *providers.FallbackChain
+	fallback         *providers.FallbackChain
-	channelManager *channels.Manager
+	channelManager   *channels.Manager
-	providerCache  map[string]providers.LLMProvider
+	providerCache    map[string]providers.LLMProvider
+	planStartPending bool // set by /plan start to trigger LLM execution
 }
 
 // processOptions configures how a message is processed
@@ -190,6 +191,23 @@ func (al *AgentLoop) Run(ctx context.Context) error {
 					SkipPlaceholder: true,
 				})
 			}
+			// /plan start sets the flag — enqueue a synthetic message so
+			// the LLM worker actually begins executing the plan.
+			if al.planStartPending {
+				al.planStartPending = false
+				select {
+				case llmQueue <- bus.InboundMessage{
+					Channel:    msg.Channel,
+					ChatID:     msg.ChatID,
+					SenderID:   msg.SenderID,
+					SessionKey: msg.SessionKey,
+					Content:    "The plan has been approved. Begin executing.",
+					Metadata:   msg.Metadata,
+				}:
+				case <-ctx.Done():
+					return nil
+				}
+			}
 			continue
 		}
 
@@ -1575,6 +1593,7 @@ func (al *AgentLoop) handlePlanCommand(args []string) (string, bool) {
 		if err := agent.ContextBuilder.SetPlanStatus("executing"); err != nil {
 			return fmt.Sprintf("Error: %v", err), true
 		}
+		al.planStartPending = true
 		return "Plan approved. Executing.", true
 
 	case "next":

pkg/agent/loop_test.go

@@ -1041,6 +1041,11 @@ func TestPlanCommand_Start(t *testing.T) {
 	if status := agent.ContextBuilder.GetPlanStatus(); status != "executing" {
 		t.Errorf("expected 'executing', got %q", status)
 	}
+
+	// planStartPending must be set so Run() enqueues an LLM trigger
+	if !al.planStartPending {
+		t.Error("expected planStartPending to be true after /plan start")
+	}
 }
 
 func TestPlanCommand_StartFromReview(t *testing.T) {
@@ -1062,6 +1067,10 @@ func TestPlanCommand_StartFromReview(t *testing.T) {
 	if status := agent.ContextBuilder.GetPlanStatus(); status != "executing" {
 		t.Errorf("expected 'executing', got %q", status)
 	}
+
+	if !al.planStartPending {
+		t.Error("expected planStartPending to be true after /plan start from review")
+	}
 }
 
 func TestPlanCommand_StartNoPhases(t *testing.T) {
@@ -1081,6 +1090,10 @@ func TestPlanCommand_StartNoPhases(t *testing.T) {
 	if status := agent.ContextBuilder.GetPlanStatus(); status != "interviewing" {
 		t.Errorf("expected status to remain 'interviewing', got %q", status)
 	}
+
+	if al.planStartPending {
+		t.Error("planStartPending must not be set when start is rejected (no phases)")
+	}
 }
 
 func TestPlanCommand_StartAlreadyExecuting(t *testing.T) {
@@ -1094,11 +1107,18 @@ func TestPlanCommand_StartAlreadyExecuting(t *testing.T) {
 	_ = agent.ContextBuilder.WriteMemory(plan)
 	al.handleCommand(context.Background(), bus.InboundMessage{Content: "/plan start"})
 
-	// Try start again
+	// Clear the flag from the first call (simulating Run() consuming it)
+	al.planStartPending = false
+
+	// Try start again — should be rejected
 	response, _ := al.handleCommand(context.Background(), bus.InboundMessage{Content: "/plan start"})
 	if !strings.Contains(response, "already executing") {
 		t.Errorf("expected 'already executing', got %q", response)
 	}
+
+	if al.planStartPending {
+		t.Error("planStartPending must not be set when plan is already executing")
+	}
 }
 
 func TestPlanCommand_Done(t *testing.T) {