fix(agent): add panic recovery to prevent zombie state after agent loop crash (#39)

Root cause: runLoop() had no recover() — panics killed the goroutine while
leaving the session queue's activeRuns entry orphaned. Subsequent messages
to the same session would hang indefinitely (zombie state).

Fix: Two-layer panic recovery:
1. loop.go: runLoop() now returns error on panic instead of crashing goroutine,
   allowing normal cleanup flow (session queue, traces, events)
2. queue.go: executeRun() defense-in-depth recovery ensures activeRuns cleanup
   and scheduleNext() even if agent-level recovery fails

internal/agent/loop.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"fmt"
 	"log/slog"
+	"runtime"
 	"sort"
 	"sync"
 	"time"
@@ -31,7 +32,18 @@ type indexedResult struct {
 	spanStart    time.Time
 }
 
-func (l *Loop) runLoop(ctx context.Context, req RunRequest) (*RunResult, error) {
+func (l *Loop) runLoop(ctx context.Context, req RunRequest) (result *RunResult, err error) {
+	defer func() {
+		if r := recover(); r != nil {
+			buf := make([]byte, 8192)
+			n := runtime.Stack(buf, false)
+			slog.Error("agent loop panicked", "agent", l.id, "session", req.SessionKey,
+				"panic", fmt.Sprint(r), "stack", string(buf[:n]))
+			result = nil
+			err = fmt.Errorf("agent loop panic: %v", r)
+		}
+	}()
+
 	// Per-run emit wrapper: enriches every AgentEvent with delegation + routing context.
 	emitRun := func(event AgentEvent) {
 		event.RunKind = req.RunKind

internal/scheduler/queue.go

@@ -2,6 +2,7 @@ package scheduler
 
 import (
 	"context"
+	"fmt"
 	"log/slog"
 	"sync"
 	"time"
@@ -294,6 +295,23 @@ func (sq *SessionQueue) startOne(ctx context.Context) {
 
 // executeRun runs the agent and then starts the next queued message(s) if capacity allows.
 func (sq *SessionQueue) executeRun(ctx context.Context, runID string, runGeneration uint64, pending *PendingRequest) {
+	// Defense-in-depth: if runFn panics despite agent-level recovery,
+	// ensure cleanup still runs so the session queue doesn't orphan this run.
+	defer func() {
+		if r := recover(); r != nil {
+			slog.Error("scheduler: executeRun panicked", "run_id", runID, "panic", fmt.Sprint(r))
+			pending.ResultCh <- RunOutcome{Err: fmt.Errorf("run panic: %v", r)}
+			close(pending.ResultCh)
+			sq.mu.Lock()
+			delete(sq.activeRuns, runID)
+			sq.removeFromOrder(runID)
+			if sq.hasCapacity() && len(sq.queue) > 0 {
+				sq.scheduleNext(sq.parentCtx)
+			}
+			sq.mu.Unlock()
+		}
+	}()
+
 	result, err := sq.runFn(ctx, pending.Req)
 	pending.ResultCh <- RunOutcome{Result: result, Err: err}
 	close(pending.ResultCh)