The cmdlet calls the /cloudinit/dump endpoint which returns YAML
content, not a task UPID. This gets passed to WaitForTask causing a
501 error. PVE regenerates cloud-init automatically on VM start, so
the explicit regenerate call is not needed for test VM setup.
Bug to fix: RegenerateCloudInitImage should PUT to /cloudinit instead
of GET from /cloudinit/dump.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The Debian .qcow2 extension was rejected by PVE's download-url API.
Ubuntu's noble-server-cloudimg-amd64.img has a .img extension that
PVE accepts, and cloud-init + qemu-guest-agent work the same way.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace raw qm/SSH commands with module cmdlets where possible:
- Invoke-PveStorageDownload for cloud image download
- New-PveVm for VM creation
- Set-PveCloudInitConfig for cloud-init user/password/IP
- Invoke-PveCloudInitRegenerate to apply config
- Start-PveVm to boot the VM
SSH/SCP retained only for operations without API support:
- Snippet upload (no snippet API)
- pvesm set for snippets content type
- qm importdisk (no importdisk cmdlet)
- qm set for --scsi0, --cicustom, --agent, --boot (not in Set-PveVmConfig)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace virt-customize with a cloud-init custom user-data snippet that
installs qemu-guest-agent on first boot. The snippet is SCP'd to the
nested PVE's /var/lib/vz/snippets/ and referenced via --cicustom.
This eliminates the need for libguestfs-tools entirely — no install
on nested PVE, no extra Docker image, no large dependencies.
Timeout increased to 300s to allow for cloud-init package installation.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- prepare-test-vm.sh: install libguestfs-tools if not present on the
fresh nested PVE before running virt-customize
- Mask PVE_PASSWORD in logs via ::add-mask:: to prevent leaking
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move virt-customize to run on the nested PVE host (which already has
libguestfs) instead of inside the CI container. This removes
libguestfs-tools, linux-image-generic, and qemu-utils from the
Dockerfile, shrinking the image from ~1.2GB to ~770MB.
- prepare-test-vm.sh: downloads Debian cloud image on nested PVE,
runs virt-customize there to install qemu-guest-agent, then creates
VM with imported disk + cloud-init
- Remove build-test-image.sh (no longer needed)
- Remove /opt/pve-images volume mount
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Infrastructure:
- build-test-image.sh: Downloads Alpine cloud image, uses virt-customize
to install qemu-guest-agent, caches at /opt/pve-images/
- prepare-test-vm.sh: SCPs image to nested PVE, creates VM via
qm importdisk with cloud-init and agent enabled, waits for agent
Workflow:
- Build Alpine image step (cached on runner, only built once)
- Deploy test VM step after provisioning, passes LINUX_VMID to tests
Tests:
- Guest Agent VM lifecycle: verify running, ACPI restart, ACPI stop
- Templates: convert Linux VM to template, clone from template
(New-PveTemplate, New-PveVmFromTemplate)
- Fix Get-PveTask: use .Upid property instead of whole PveTask object
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
PVE 8 (Bookworm) may not have /etc/apt/sources.list.d/debian.sources,
causing the grep fallback to use 'trixie' which is invalid for PVE 8.
Use /etc/os-release VERSION_CODENAME which works on both PVE 8 and 9.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
/tmp on the self-hosted runner has sticky bit — files created by
previous runs (or different users) can't be overwritten. RUNNER_TEMP
is per-run and owned by the runner user.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The proxmox-auto-install-assistant defaults --tmp to the source ISO's
directory (/opt/pve-isos/), which the runner user can't write to.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- first-boot.sh: Remove .sources files (PVE 9 DEB822 format) not just
.list, and detect suite (trixie vs bookworm) from debian.sources
- create-api-token.sh: Accept full endpoint URL (fixes missing port),
use /access/domains for API check (no auth needed on PVE 9),
avoid pipefail on curl|python3 pipe
- preflight-cleanup.sh: New script that runs at start and end of every
provisioning run — cleans orphaned VMs, ISOs, and stale terraform state
- Workflow: Add pre-flight and final cleanup steps so failed runs can't
leave the runner/host in a broken state
All changes validated manually on the self-hosted runner.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Use proxmox_virtual_environment_file resource to upload the auto-install
ISO via the API. terraform destroy now cleans up both the VM and ISO.
- Remove manual upload-to-pve.sh and cleanup-pve-storage.sh scripts.
- Commit .terraform.lock.hcl for reproducible CI builds.
- Replace iso_file_id variable with iso_local_path and iso_storage.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Workflow now provisions a throwaway nested PVE VM via Terraform, runs
integration tests against it, then destroys it. Uses proxmox-auto-install-assistant
to bake the answer file and a first-boot script (installs qemu-guest-agent)
directly into the ISO. IP is discovered via the QEMU guest agent on the
parent PVE, eliminating the need for static IP configuration.
Supports both PVE 8.x and 9.x via workflow_dispatch version selector.
Skip provisioning with skip_provision=true to test against a pre-existing PVE.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Provisions a throwaway nested Proxmox VE instance on an existing host
using bpg/proxmox provider. Includes unattended install via answer file,
API wait script, and automatic API token creation. Designed for CI-driven
integration test environments.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Clint Branham