* Connect to remote SSH nodes using SSH CLI
Replace Go-native SSH implementation with SSH CLI execution
to support diverse SSH configurations and agents. Implements
'uncloudd dial-stdio' subcommand that proxies gRPC connections
over stdin/stdout, similar to Docker's approach.
This change addresses compatibility issues with:
- SSH agents exposing many keys (1Password, causing "too many
authentication failures")
- Tailscale SSH (which doesn't support advanced SSH channel
types like direct-streamlocal)
- Custom SSH configurations in ~/.ssh/config
The dial-stdio approach reduces SSH feature requirements by
streaming the unix socket connection over stdin/stdout instead
of using SSH channel forwarding.
Changes:
- Add 'uncloudd dial-stdio' hidden subcommand for socket proxy
- Add SSHCLIConnector using ssh command + dial-stdio
- Update connection logic to use SSH CLI connector
- Maintain backward compatibility with SSHKeyFile config
Resolves#131
* Fix sshcli tests missing ConnectionTimeout
Introduced short connection timeout on the first change but forgot to
update tests to match.
* Add SSHCLI field and update MachineConnection String() format
- Add SSHCLI field to support ssh_cli YAML configuration
- Update String() to use URI-like format (ssh://, ssh+cli://, tcp://)
* Add Validate() method and tests for MachineConnection
- Add Validate() to ensure connection methods are mutually exclusive
- Add tests for validation and String() method
* Unify SSH connector configs to use SSHConnectorConfig
* Update connectCluster to support both SSH connector types
* Restore Go SSH connector as default for machine init/add
Revert provisionOrConnectRemoteMachine to use Go SSH connector:
- Root users: reuse SSH connection from provisioning
- Non-root users: establish new connection for group membership
- Remove SSH CLI as default connector
SSH CLI connector remains available via ssh_cli config field.
* Add sshCLIDialer with DialContext method
Implement proxy.ContextDialer for SSHCLIConnector using SSH -W flag.
Each dial spawns a new SSH process for TCP forwarding, enabling
independent connections separate from the gRPC dial-stdio connection.
* Implement SSHCLIConnector.Dialer() method
Return sshCLIDialer instead of error, enabling uc image push
functionality with SSHCLIConnector. Validates connector is
configured before returning dialer.
* Fix half-closing implementation matchin Docker's approach
* Use testify assertions for connection tests
* Allow ssh+cli:// to be used with --connect
This way I can skip the configuration file while testing things out,
and confirm it works correctly:
$ unset SSH_AUTH_SOCK
$ ./uncloud --connect ssh://provision@blatta11 machine ls
Error: connect to cluster: connect to machine: SSH login to
provision@blatta11:22: connect using SSH agent: connect to SSH
agent: dial unix: missing address
$ ./uncloud --connect ssh+cli://provision@blatta11 machine ls
NAME STATE ADDRESS PUBLIC IP WIREGUARD ENDPOINTS MACHINE ID
blatta11 Up 10.210.0.1/24 - 100.64.0.22:51820, ...
* Validates configuration before connecting to cluster
* Do not tie client constructor with real validation
No longer attempt to validate the connection when instantiating a new
client.
Later on we could validate it in different places.
* Cleanup test and remove AI-slop
There were some serious slop in those tests, so took the time to clean
them up and kept only the relevant ones.
There is some repetition between buildSSHArgs and buildDialArgs but can
be tackled at a later stage.
* Fix connection representation tests
Prefix connection with ssh ssh+cli respectively.
* Wait for stdout before returning
Missed copy & pasta from Docker dial-stdio implementation (this happens
when you stare at the code for too long that it burns your eyes).
* feat: Adding a "nearest.service.local" mode to internal DNS
* Pass listenAddr to internal DNS server (rather than calling into network.MachineIP)
Also adds an exact round-robin via "rr.{service}" mode. Not sure we want that over random or not, though.
* Remove the resolved count and rotating round-robin. "rr.name" mode is now the same as default with random shuffle.
* Add test on nearest internal DNS lookup
* Add docs on internal DNS order modes
* feat: support compose.yaml `service.gpus` and `service.deploy.resources.reservations.devices` DeviceRequests to Docker container
* Fix lint error
* Adjust test comments to work with linter
* Rename DeviceRequests to DeviceReservations to be more consistent with compose nomenclature
* Device reservation changes are immutable and should trigger a container recreate rather than update
* Set machine id in service container ENV
* Add machine id to clone of the container's env map to avoid exposing serialization format
* fix: panic on setting UNCLOUD_MACHINE_ID env var when Env is nil in the spec
---------
Co-authored-by: Pasha Sviderski <me@psviderski.name>
Pasha Sviderski