[ideaship]

ssh_args enables connection multiplexing but no server-alive probes. When the transport under an ansible session dies without the client being told, the play waits until the kernel's TCP keepalive notices, which is 7200 s by default in the osism-ansible container.

It bites when bootstrap's package tasks upgrade systemd, or a library its daemons link against, so that the upgrade re-executes systemd and restarts systemd-networkd while the task is running. One such update is enough: nodes installed from an image built minutes before the update was published are affected too, and the likelihood grows with the image's age. Killing the wedged ssh -o ControlMaster clients in osism-ansible unblocks the play immediately, and the tasks complete normally.

This PR adds -o ServerAliveInterval=30 -o ServerAliveCountMax=10. The client then drops a session whose peer has not answered for 5 minutes. That is what the manual recovery does, but bounded and automatic.

• sshd answers the probes at protocol level even while a command runs silently, so long quiet tasks (tempest, large upgrades) are not affected. Only a dead transport, or a host whose sshd has not answered for 5 minutes, is dropped.
• Reboots are not affected. The reboot playbook fires shutdown asynchronously and reconnects with fresh connections (wait_for_connection or ansible.builtin.reboot), so no session has to survive a reboot.
• The values are generous (5 minutes rather than, for example, 45 s) to leave margin for loaded hosts and lossy links.
• Operators can still override ssh_args in environments/ansible.cfg, which the inventory reconciler merges over this file.

🤖 Generated with Claude Code