NVIDIA / holodeck / 29150898618

52%

main: 51%

push

github

fix(ssh): reject sshConfig in cluster mode (#851 T8b)

Cluster-mode SSH semantics (per-node bastion, agent auth, host-key
policy, timeouts) are undesigned. Rather than leave auth.sshConfig
silently inert on a cluster-mode env (no bastion, no strict policy, no
timeouts applied), fail loudly at every touchpoint T8 already wired:

- EnvironmentSpec.ValidateSSHConfigMode (api/holodeck/v1alpha1) is the
  single source of truth: cluster-mode + sshConfig set -> reject with
  'auth.sshConfig is not yet supported in cluster mode (see
  NVIDIA/holodeck#851); remove the sshConfig block or use single-node
  mode'. Called from the same three ingestion points as T8's
  SSHConfig.Validate(): create's Before hook, dryrun's Before hook, and
  the CI entrypoint — all before any cloud/SSH action.
- pkg/provisioner.NewClusterProvisioner captures the rejection in an
  unexported sticky-error field (its exported signature can't change
  without breaking existing callers/tests); ProvisionCluster and
  GetClusterHealth surface it before touching any node — defense in
  depth mirroring T8's provisioner.New choke point.

Also documents the gap on the SSHConfig type comment (single-node only)
and drafts (not filed) a follow-up issue at
.superpowers/sdd/issue-draft-cluster-sshconfig.md covering the design
questions for wiring sshConfig through the 7 in-package New() call
sites in pkg/provisioner/cluster.go.

Signed-off-by: Carlos Eduardo Arango Gutierrez <eduardoa@nvidia.com>

22 of 25 new or added lines in 5 files covered. (88.0%)

6833 of 13171 relevant lines covered (51.88%)

0.58 hits per line