fix: don't reconnect after a command longer than the ack timeout

The daemon handles each message inline, so while a command runs its serve
loop reads nothing and the controller's heartbeat acks sit unread in the
socket. When the command finished, the next ack_check saw no ack for longer
than ack_timeout (30s by default) and reconnected, so the daemon vanished
from the registry for its 5s backoff right after answering.

This broke the E2E run on #44: `microdnf install -y curl` on rocky took
the full 60s timeout, daemon-rocky-1 reconnected the same instant, and the
following tests failed with "Daemon daemon-rocky-1 not found".

Restart the ack deadline after each inline message. The buffered acks are
read on the next iterations; a controller that really stopped acking is
still caught one ack_timeout later. Introduced in #40.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: kerthcet <kerthcet@gmail.com>

Author: kerthcet

python/tests/test_e2e.py

@@ -278,6 +278,23 @@ def test_daemon_restart(self, server):
         assert result.success
         assert "after" in result.stdout
 
+    def test_daemon_stays_connected_after_long_command(self, server):
+        """A command longer than the daemon's ack timeout (30s) must not make it reconnect
+
+        The daemon handles a command inline, so heartbeat acks go unread while it runs.
+        It used to read that as an unresponsive controller and reconnect right after
+        answering, so the next call failed with "daemon not found".
+        """
+        result = server.exec("daemon-debian-2", "sleep 35 && echo done", timeout=60)
+        assert result.success
+        assert "done" in result.stdout
+
+        # No wait: a reconnecting daemon is missing from the registry for its 5s backoff.
+        time.sleep(2)
+        result = server.exec("daemon-debian-2", "echo still-here", timeout=5)
+        assert result.success
+        assert "still-here" in result.stdout
+
 
 class TestE2EStats:
     """Test statistics with Docker daemons"""

sandd/src/main.rs

@@ -620,6 +620,15 @@ where
                 {
                     error!("Error handling message: {}", e);
                 }
+
+                // While the message was handled inline (an exec can run for minutes)
+                // this loop read nothing, so the controller's acks sat unread in the
+                // socket. That is OUR silence, not the controller's: without this, any
+                // command longer than ack_timeout made the next ack_check reconnect,
+                // dropping the daemon from the registry right after it answered. The
+                // buffered acks are read on the next iterations; a controller that
+                // really stopped acking is still caught one ack_timeout from now.
+                last_ack = Instant::now();
             }
         }
     };