fix: Fixes to server shutdown stream

agent/internal/sync/sync.go

@@ -118,9 +118,21 @@ func poll(client *grpcclient.Client, cfg *config.Config, version string) error {
 	return nil
 }
 
+// How long a command stream must survive before it counts as having worked.
+// Past this, the next drop is treated as a fresh incident rather than as the
+// continuation of a run of failures.
+const streamHealthyAfter = time.Minute
+
 func runCommandStream(ctx context.Context, cfg *config.Config) {
 	backoff := time.Second
-	const maxBackoff = 2 * time.Minute
+
+	// Two minutes was the old ceiling, and it was reached far too easily. The
+	// command stream is what makes this agent controllable at all: while it is
+	// down, workflows and console sessions fail as "agent offline" even though
+	// SyncKeys keeps polling happily and the fleet list still shows the server
+	// active. A shorter ceiling costs a few reconnect attempts; the old one cost
+	// two minutes of an agent that looks fine and answers nothing.
+	const maxBackoff = 30 * time.Second
 
 	for {
 		select {
@@ -129,22 +141,40 @@ func runCommandStream(ctx context.Context, cfg *config.Config) {
 		default:
 		}
 
-		if err := connectAndHandleStream(ctx, cfg); err != nil {
-			if ctx.Err() != nil {
-				return
-			}
-			log.Printf("command stream error: %v, reconnecting in %s", err, backoff)
-			select {
-			case <-ctx.Done():
-				return
-			case <-time.After(backoff):
-			}
-			if backoff < maxBackoff {
-				backoff *= 2
-			}
-		} else {
+		started := time.Now()
+		err := connectAndHandleStream(ctx, cfg)
+		if ctx.Err() != nil {
+			return
+		}
+
+		// A stream that stayed up is evidence the control plane is reachable,
+		// whatever ended it. Without this the backoff only ever climbed:
+		// connectAndHandleStream returns an error on *every* stream end,
+		// including a healthy one dropped by a routine deploy, so an agent
+		// pinned itself at the ceiling after a handful of ordinary restarts and
+		// stayed there for the rest of its life.
+		if time.Since(started) >= streamHealthyAfter {
 			backoff = time.Second
 		}
+
+		if err != nil {
+			log.Printf("command stream error: %v, reconnecting in %s", err, backoff)
+		} else {
+			log.Printf("command stream closed, reconnecting in %s", backoff)
+		}
+
+		select {
+		case <-ctx.Done():
+			return
+		case <-time.After(backoff):
+		}
+
+		if backoff < maxBackoff {
+			backoff *= 2
+			if backoff > maxBackoff {
+				backoff = maxBackoff
+			}
+		}
 	}
 }