Running Stabilization Loop

Running Stabilization Loop

Overview

Use this workflow when the goal is not only to detect failures, but to fix them and prove full regression safety after major changes.

Usage

/running-stabilization-loop

Buddy Prompt Template

Use this exact prompt when delegating the loop to another agent:

Run the full kube-hetzner v3 stabilization loop end-to-end, do not stop at pass/fail reporting.

Goals:
1) Make every matrix scenario pass from simple to complex.
2) Fix root causes in kube-hetzner (not workarounds in test harness) unless explicitly required.
3) Re-run full matrix after fixes to prove no regressions.
4) Validate upgrade path from latest git tag to the current release-candidate branch for both k3s and rke2.
5) When Tailscale credentials are available, include the Tailscale node-transport
   scenarios: secure single-network Tailnet access, plus Flannel with multiple
   Hetzner Networks, route auto-approval, autoscaler scale-up/down,
   `tailscale status`, and cross-network pod traffic.
6) Include final v3 polish scenarios when touched: Cilium Gateway API with
   `enable_kube_proxy=false`, embedded registry mirror on k3s or RKE2, endpoint
   output sanity, and example validation.

Hard requirements:
- Use /Users/karim/.ssh/id_ed25519 with IdentitiesOnly=yes for all SSH.
- Use /Volumes/MysticalTech/Code/kube-test/run_v3_matrix.sh and summary-*.md as source of truth.
- If RKE2/K3s apply hangs, SSH during apply and inspect journalctl/cloud-init immediately.
- If destroy hangs, run cleanup and delete stuck autoscaled servers via hcloud.
- Keep changes minimal and production-safe; run Terraform and OpenTofu validation before each rerun.
- For Tailscale node transport, verify single-network clusters work without
  route approval when route advertisement is disabled; for multinetwork, verify
  node-private routes are approved and SNAT is disabled for advertised routes.
  Do not mistake external overlay SSH access for supported Kubernetes node
  transport.
- For Cilium Gateway API, verify GatewayClass exists, Gateway is accepted, and
  HTTPRoute serves through the Hetzner LoadBalancer. For embedded registry
  mirror, verify `registries.yaml`, `embedded-registry: true`, peer ports, and
  image-pull behavior.

Deliverables:
- Final scenario matrix results with log paths.
- List of fixes made with files changed.
- Upgrade test result (tag -> release-candidate branch) and any required migration notes.
- Confirmation cleanup is complete and no khv3-* resources are left.

Test Environment

• Module code: /Volumes/MysticalTech/Code/kube-hetzner
• Test harness: /Volumes/MysticalTech/Code/kube-test
• Matrix runner: /Volumes/MysticalTech/Code/kube-test/run_v3_matrix.sh
• SSH key (required): /Users/karim/.ssh/id_ed25519

Invariants

1. Always use -o IdentitiesOnly=yes -i /Users/karim/.ssh/id_ed25519 for SSH.
2. Always start from a clean scenario state before new loops.
3. If any matrix scenario fails, debug and fix it before continuing.
4. After targeted fixes, rerun the full matrix to guard against regressions.
5. For destroy hangs with autoscaler/LB leftovers, use cleanup and hcloud deletes.
6. Do not keep temporary broad firewall overrides after root cause is fixed.
7. Treat scenarios_v3/summary-*.md + scenario logs/*.log as canonical run evidence.

Workflow

digraph stabilize_flow {
    rankdir=TB;
    node [shape=box];

    sync [label="1. Sync branch + terraform gates"];
    clean [label="2. Kill stale runs + cleanup state"];
    images [label="3. Verify/build required snapshots"];
    matrix [label="4. Run matrix (simple -> complex)"];
    debug [label="5. Proactive SSH/journal triage"];
    fix [label="6. Implement fix + rerun failing scenarios"];
    full [label="7. Full matrix regression rerun"];
    upgrade [label="8. Upgrade path tests (latest tag -> release candidate)"];
    finalize [label="9. Final cleanup + commit/push"];

    sync -> clean;
    clean -> images;
    images -> matrix;
    matrix -> debug;
    debug -> fix;
    fix -> full;
    full -> upgrade;
    upgrade -> finalize;
}

Step 1: Sync Branch and Run Gates

cd /Volumes/MysticalTech/Code/kube-hetzner
git pull --ff-only origin "$(git branch --show-current)"
terraform fmt -recursive
terraform init -backend=false
terraform validate
tmpdir="$(mktemp -d)"
rsync -a --exclude .git --exclude .terraform --exclude .terraform-tofu ./ "$tmpdir"/
(cd "$tmpdir" && tofu init -backend=false && tofu validate)
rm -rf "$tmpdir"

cd /Volumes/MysticalTech/Code/kube-test
terraform init -upgrade

Step 2: Kill Stale Runs and Clean State

pkill -f run_v3_matrix.sh || true
pkill -f "terraform -chdir=/Volumes/MysticalTech/Code/kube-test/scenarios_v3" || true

Destroy all scenarios:

for d in \
  s1-k3s-basic \
  s2-rke2-basic \
  s3-k3s-autoscaler \
  s4-rke2-autoscaler \
  s5-leapmultiarch \
  s6-mixed-os \
  s7-kubeapi-join-lb
do
  terraform -chdir="/Volumes/MysticalTech/Code/kube-test/scenarios_v3/$d" destroy -auto-approve || true
done
find /Volumes/MysticalTech/Code/kube-test/scenarios_v3 -name ".terraform.tfstate.lock.info" -delete

If resources are stuck (especially autoscaler nodes), run:

cd /Volumes/MysticalTech/Code/kube-test
/Volumes/MysticalTech/Code/kube-hetzner/scripts/cleanup.sh --dry-run
/Volumes/MysticalTech/Code/kube-hetzner/scripts/cleanup.sh --execute

The force cleanup treats the Terraform token's entire HCloud project as cluster-dedicated. Inspect every dry-run row before the typed confirmation and include volumes, snapshots, DNS zones, or Storage Boxes only deliberately.

Optional alias:

alias cleanupkh='/Volumes/MysticalTech/Code/kube-hetzner/scripts/cleanup.sh'

Step 3: Verify Snapshot Prerequisites

Expected snapshots:

• Leap Micro: k3s x86, k3s arm, rke2 x86, rke2 arm (4 total)
• MicroOS: k3s x86, k3s arm, rke2 x86, rke2 arm (4 total, required when the release touches the legacy image path or mixed-os scenarios)

Check:

hcloud image list --type snapshot -o columns=id,name,architecture,labels

Create missing ones from a fresh, manifest-verified generated bundle. Keep the token already exported in the environment; do not parse or print it:

repo_root=/Volumes/MysticalTech/Code/kube-hetzner
build_root="$(mktemp -d)"
KH_SOURCE_DIRECTORY="$repo_root" folder_name=generated \
  folder_path="$build_root" create_snapshots=none \
  "$repo_root/scripts/create.sh"
cd "$build_root/generated/packer"
./scripts/install-verified-packer-plugin-hcloud.sh
for template in hcloud-leapmicro-snapshots.pkr.hcl hcloud-microos-snapshots.pkr.hcl; do
  packer init "$template"
  for distro in k3s rke2; do
    packer build -var "selinux_package_to_install=${distro}" "$template"
  done
done

For a release canary, set KH_SOURCE_DIRECTORY to the exact release checkout or an extracted archive of that release. Record the generated bundle's manifest digest, snapshot IDs, architectures, and distro labels before any cluster apply. Never build from loose templates copied into kube-test.

Step 4: Run Matrix (Simple to Complex)

Run full matrix:

cd /Volumes/MysticalTech/Code/kube-test
RUN_ID="$(date +%Y%m%d-%H%M%S)" ./run_v3_matrix.sh | tee "scenarios_v3/matrix-${RUN_ID}.log"

Run targeted scenarios:

cd /Volumes/MysticalTech/Code/kube-test
RUN_ID="$(date +%Y%m%d-%H%M%S)" SCENARIOS="s2-rke2-basic s4-rke2-autoscaler" ./run_v3_matrix.sh

Step 5: Proactive Debugging During Runs

Do not wait for timeout-only diagnosis. SSH while apply is running if progress stalls.

Find nodes:

hcloud server list -o columns=id,name,status,ipv4 | grep khv3-

SSH:

ssh -i /Users/karim/.ssh/id_ed25519 \
  -o IdentitiesOnly=yes \
  -o StrictHostKeyChecking=no \
  -o UserKnownHostsFile=/dev/null \
  root@<node-ip>

Core checks:

systemctl is-system-running
cloud-init status --long || true
journalctl -u cloud-init -u cloud-config -u sshd --no-pager -n 200
journalctl -u k3s -u k3s-agent -u rke2-server -u rke2-agent --no-pager -n 300

RKE2-specific checks:

ls -lah /etc/rancher/rke2
cat /etc/rancher/rke2/config.yaml
systemctl cat rke2-server
systemctl cat rke2-agent

If terraform-installed scripts fail, run the exact failing commands manually and capture output (do not only read terraform stderr).

Known high-frequency failures to check first:

1. Root SSH lock on Leap Micro:
   • Symptom: User root not allowed because account is locked
   • Check: grep '^root:' /etc/shadow
   • Fix path: packer transactional-update stage, not volatile one-off edits.
2. Firewall/API source mismatch false negatives:
   • Symptom: cluster healthy internally but kubectl/API unreachable from runner.
   • Check runner public source vs firewall_kube_api_source.
   • Fix: use explicit allowed source in test-only scenarios, then revert broad overrides.
3. RKE2 config/path issues:
   • Confirm module writes persistent config under /etc/rancher/rke2.
   • If a script writes to non-persistent/incorrect paths, fix module scripts and rerun.
4. Stale Leap Micro signing key:
   • Symptom: transactional-update.service reports NOTTRUSTED packages or repository metadata signed by an expired 35A2F86E29B700A4 key.
   • Check: rpm -q openSUSE-build-key gpg-pubkey-29b700a4-6a17fa38.
   • Fix path: rebuild the Leap Micro snapshot from the current authenticated openSUSE appliance. Never use --no-gpg-checks, gpgcheck=0, rpm --nosignature, or an equivalent bypass.

Leap Micro OS update release gate

For release candidates that touch snapshots, bootstrap, SSH, SELinux, or OS updates, keep a disposable cluster after apply and prove the real update path. Run the agent first, then a control plane:

transactional-update --no-selfupdate --non-interactive up

Require exit status 0, a new default snapshot when updates are available, and no NOTTRUSTED, key-import, RPM-lock, or discarded-snapshot failure. Reboot the node and verify all of the following before continuing:

uname -r
rpm -q openSUSE-build-key gpg-pubkey-29b700a4-6a17fa38
systemctl is-active k3s k3s-agent rke2-server rke2-agent 2>/dev/null || true
systemctl --failed
getenforce
sshd -T | grep -E '^(usepam|permitrootlogin|pubkeyauthentication|passwordauthentication|kbdinteractiveauthentication) '
snapper list

Also prove that the boot ID changed, the node returned Ready, all required pods recovered, API /readyz passes, metrics work, and an in-cluster DNS probe succeeds. A single-control-plane canary should show brief API downtime; require full recovery before destroying it. Re-run the checks after every reboot rather than accepting a successful package transaction as deployment proof.

Step 6: Fix + Targeted Rerun Loop

1. Implement minimal fix in /Volumes/MysticalTech/Code/kube-hetzner.
2. Run gates:

cd /Volumes/MysticalTech/Code/kube-hetzner
terraform fmt -recursive
terraform init -backend=false
terraform validate
tmpdir="$(mktemp -d)"
rsync -a --exclude .git --exclude .terraform --exclude .terraform-tofu ./ "$tmpdir"/
(cd "$tmpdir" && tofu init -backend=false && tofu validate)
rm -rf "$tmpdir"

3. Rerun only failed scenarios:

cd /Volumes/MysticalTech/Code/kube-test
RUN_ID="$(date +%Y%m%d-%H%M%S)" SCENARIOS="<failed-scenarios>" ./run_v3_matrix.sh

Repeat until targeted failures are resolved.

Step 7: Full Regression Rerun

After targeted failures are fixed, rerun all scenarios:

cd /Volumes/MysticalTech/Code/kube-test
RUN_ID="$(date +%Y%m%d-%H%M%S)" ./run_v3_matrix.sh

No change is done until this full rerun passes.

Step 8: Upgrade Path Loop (Latest Tag -> Release Candidate)

Build a baseline on latest tag, then upgrade to the current release-candidate code.

LATEST_TAG="$(git -C /Volumes/MysticalTech/Code/kube-hetzner tag --list 'v*' --sort=-v:refname | head -n1)"
TAG_TREE="/tmp/kube-hetzner-${LATEST_TAG}"
git -C /Volumes/MysticalTech/Code/kube-hetzner worktree add --detach "$TAG_TREE" "$LATEST_TAG"

Create an upgrade scenario from a known-good matrix scenario (repeat for both k3s and rke2):

UPG="/Volumes/MysticalTech/Code/kube-test/scenarios_v3/upgrade-k3s-${LATEST_TAG#v}"
cp -R /Volumes/MysticalTech/Code/kube-test/scenarios_v3/s1-k3s-basic "$UPG"
sed -i '' "s|source    = \"/Volumes/MysticalTech/Code/kube-hetzner\"|source    = \"$TAG_TREE\"|" "$UPG/main.tf"
terraform -chdir="$UPG" init -upgrade
terraform -chdir="$UPG" apply -auto-approve

sed -i '' "s|source    = \"$TAG_TREE\"|source    = \"/Volumes/MysticalTech/Code/kube-hetzner\"|" "$UPG/main.tf"
terraform -chdir="$UPG" init -upgrade
terraform -chdir="$UPG" plan
terraform -chdir="$UPG" apply -auto-approve
terraform -chdir="$UPG" destroy -auto-approve

Cleanup worktree:

git -C /Volumes/MysticalTech/Code/kube-hetzner worktree remove "$TAG_TREE"

Step 9: Finalize

If requested to push the current release-candidate branch:

cd /Volumes/MysticalTech/Code/kube-hetzner
git status --short
git add -A
git commit -m "fix: stabilize v3 matrix and upgrade loops"
git push origin "$(git branch --show-current)"

Abort/Cleanup Shortcut

pkill -f run_v3_matrix.sh || true
pkill -f "terraform -chdir=/Volumes/MysticalTech/Code/kube-test/scenarios_v3" || true
for d in /Volumes/MysticalTech/Code/kube-test/scenarios_v3/s*; do
  terraform -chdir="$d" destroy -auto-approve || true
done
find /Volumes/MysticalTech/Code/kube-test/scenarios_v3 -name ".terraform.tfstate.lock.info" -delete