pia_forward_bypass_chains held up against real production conditions,
not just a short recheck: 1528 healthy passes, 0 SAFETY DRIFT events,
0 safety_drift classifications since the fix was deployed
(2026-09-22 20:37 -> 2026-09-23 22:02). kube-router kept resyncing
throughout; the bypass tolerance handled it silently every time.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found live, 2026-09-22, minutes after first enabling
pia-gateway-recover.timer (rollout step 3): a real, ~10-minute outage
loop. pia-gateway-recover and pia-killswitch-reassert are two
independently-scheduled 60s timers with no coordination. kube-router's
actual resync cadence turned out to be ~60-90s on this cluster - far
more frequent than the "boot-scale, rare" assumption behind the
previous reassert-timer fix - so a recover pass would periodically
sample iptables -S FORWARD in the window between a kube-router
push-down and the next reassert fix, correctly detect real (if
momentary) drift, and call emergency_close(), which unconditionally
forces table pia's route to unreachable AND resets
consecutive_failures to 0 in the same step. Journal timeline:
19:38:58 3rd unhealthy check -> restart
19:39:09 recovered - route open, handshake fresh
19:40:03 drift again -> emergency close -> route forced closed,
counter reset to 0
19:41:08 drift again -> emergency close again
VLAN 50 was open ~10s out of every 60-90s cycle. Confirmed via the
alert-bridge Discord notifications (working end to end, first real
trigger) and the full pia-gateway-recover/pia-killswitch-reassert
journals. Mitigated live: stopped pia-gateway-recover.timer, manually
restarted wg-quick@pia-wg to reopen the route.
This wasn't a detection bug - the drift was real. It was that treating
every momentary reordering as a full safety emergency was stricter
than the actual invariant required, given kube-router's own resync
cadence on this host - and we'd already proven (reading
KUBE-ROUTER-FORWARD's actual live content) that this specific,
recurring reordering cannot affect VLAN 50 traffic regardless of
position, since every rule in it is scoped to specific pod /32
addresses.
Adds pia_forward_bypass_chains (Classifier
._forward_jump_effectively_first): the FORWARD-jump safety check now
tolerates a configured, named chain sitting ahead of our jump, but
only for as long as that chain's OWN current content re-verifies -
live, on every single check, never a one-time trust - as structurally
unable to match 10.10.50.0/24 (every -s/-d value on every rule
disjoint from our subnet; a bare/unconditional rule, a negated match,
or any construct not explicitly recognized fails closed). Scoped to
KUBE-ROUTER-FORWARD only, not DOCKER-USER/DOCKER-FORWARD - dockerd
sets its hooks up once at daemon start, not on a recurring cycle, and
DOCKER-FORWARD itself jumps to three further nested chains never
fully content-verified, so it stays a hard drift trigger. An
unrecognized chain, or a bypass-listed chain whose content stops
verifying safe, still fails closed and emergency-closes exactly as
before.
Added 11 tests (105 total, all passing): the exact live topology
tolerated when bypass-configured and still blocked by default,
duplicate kube-router jumps (the exact live anomaly) tolerated, a
bypass chain picking up a matching source/bare rule/negated match
still blocking, an unrecognized chain still blocking even with a
bypass list configured, a FORWARD-level rule with a real matcher never
treated as a bypass candidate, a missing bypass chain blocking, and
two end-to-end cmd_recover tests confirming no emergency close on the
tolerated case and a real emergency close still firing on a genuine
drift.
pia-gateway-recover.timer is currently stopped on minisforum pending
redeployment and live re-verification of this fix (rollout step 3 is
not yet complete).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found live, 2026-09-22: after fixing pia-killswitch.sh's FORWARD-jump
repositioning and getting one pia-gateway register attempt past the
safety check, the very next attempt - roughly an hour later, no
reboot, no UFW reload, no script content change in between - hit the
identical "FORWARD jump to kill-switch chain is not first" failure
again. This disproves the earlier assumption that kube-router/dockerd
reassert their own position-1 jumps only at boot/restart scale:
kube-router resyncs on ordinary Kubernetes API events, and this
cluster's CronJobs (ca-sync, porkbun-ddns, gitea-backup) each
create/delete a Pod every few minutes - plausible triggers for
kube-router's own reconciliation loop, and therefore plausible
triggers for this drift, recurring on a timescale of minutes.
The reposition logic itself (previous commit) is correct - it just
only ran on the existing triggers (boot, UFW reload, an
Ansible-detected script content change), none of which fired between
the two attempts.
Adds pia-killswitch-reassert.timer/.service: re-runs the already fully
idempotent pia-killswitch.sh every pia_healthcheck_interval_sec (same
cadence as the health check), giving it regular chances to run instead
of only sporadic ones. Deliberately separate from
pia-gateway-recover.timer and NOT gated by pia_auto_recovery_enabled -
this is the kill switch continuing to assert its own already-
established, narrowly-scoped state (source rule, table default,
FORWARD jump position), the same kind of thing it already does
unconditionally at boot and after a UFW reload, just on a schedule
too - not the higher-level automatic recovery/registration logic the
plan was deliberately cautious about running unattended.
Unblocked live in the meantime with a manual
`sudo systemctl restart pia-killswitch.service`.
Offline checks: yamllint, ansible-playbook --syntax-check, rendered
templates (Jinja2 + bash -n on the updated script), ansible-lint
(diffed against main - no new finding categories), full pia_gateway.py
test suite (94/94, unaffected by this change) all pass.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found live, 2026-09-22, on the addKey step of the first successful
kill-switch-check pass: every candidate failed with
kind=tls detail="certificate verification failed". Root-caused on
minisforum: PIA's own CA certificate (files/pia-ca.crt) has a CA:TRUE
Basic Constraints extension that isn't marked critical - a real RFC
5280 non-conformance, confirmed not a stale vendored copy (a fresh
fetch from pia-foss/manual-connections is byte-identical, same
sha256 already recorded in the file's own header, same defect).
Recent OpenSSL (3.2+; confirmed via a direct repro against
minisforum's 3.5.7 - SSLCertVerificationError, verify_message
"Basic Constraints of CA cert not marked critical") rejects this by
default, a check promoted from X509_STRICT-only into the baseline
chain-verification path. openssl s_client (LibreSSL, this dev
machine) and curl both accept the identical cert/chain/CA file - only
Python's ssl module, via this host's system OpenSSL, rejects it. This
went unnoticed until now because the three prior successful
registrations all went through curl --cacert (the legacy Ansible
path); this was the first time the Python ssl code path ran for real.
Python's stdlib ssl module has no custom verify-callback hook, so
there's no clean way to tolerate just this one known,
third-party-controlled defect while keeping full validation for
everything else. Rather than weakening verify_mode or adding a new
dependency (pyOpenSSL, for a real callback - considered, rejected as
a bigger lift than necessary) to route around a gap in a library we
don't control, addKey now shells out to curl
(HttpClient.get_pinned_curl, curl added to this role's apt list) -
the same tool PIA's own reference client already uses for this exact
endpoint, with --cacert and never -k/--insecure, so this is a
different, already-proven-correct TLS backend for the one endpoint
that needs it, not a downgrade. serverlist/token are unaffected
(public WebPKI certs) and still use the stdlib ssl path.
Verified: a real subprocess curl call against the actual failing
candidate (Server-11669-10a / 69.33.198.97) gets a genuine HTTP
response from PIA, proving TLS validation succeeds via this path.
Also found and handled: curl's own exit code for a --fail-triggered
rejection isn't fully portable (macOS/SecureTransport curl reports 56
where GNU/OpenSSL curl reports 22 for the same case) - classification
now also matches curl's own error text as a second signal, not just
the exit code.
Added 8 tests (94 total, all passing) covering the curl invocation
shape, stdin token delivery (no trailing newline, never in argv), and
exit-code + text-based error classification including the
cross-platform exit-code case found live.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
_request()'s ssl.SSLCertVerificationError handler hardcoded "certificate
verification failed" regardless of cause, discarding e.verify_message/
e.reason (standard OpenSSL diagnostic text - never a secret) - exactly
what's needed to tell a wrong CA, expired cert, clock skew, and a
hostname mismatch apart. Found while diagnosing a live addKey TLS
failure on minisforum that doesn't reproduce from this machine (same
CA file, same pinned IP, same SNI hostname - openssl s_client and a
direct repro of get_pinned() both succeed here), so the next attempt
needs the real reason, not a collapsed generic string. Also stopped
discarding detail on the generic SSLError/OSError branches.
Verified: forcing a real verification failure (wrong CA) now reports
"unable to get local issuer certificate" instead of the old generic
text. Full test suite still passes (86/86).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Found live on the first deployment of the auto-recovery controller
(2026-09-22): pia-gateway register correctly refused with "FORWARD
jump to kill-switch chain is not first". Root cause confirmed on
minisforum via iptables -S FORWARD/-S KUBE-ROUTER-FORWARD/-S
DOCKER-USER/-S DOCKER-FORWARD: the PIA-VLAN50 jump was 4th, behind
kube-router's and dockerd's own jumps, which each independently
reassert position 1 on their own resyncs/restarts (kube-router had
even left a duplicate of its own rule behind). pia-killswitch.sh.j2
only checked whether its jump existed (iptables -C) before deciding
not to touch anything — existence isn't proof of position, so once
pushed down it stayed down.
Confirmed safe to reposition unconditionally by reading the actual
content of what's ahead of it rather than assuming: KUBE-ROUTER-FORWARD
only matches specific pod /32 addresses, DOCKER-USER was empty, and
DOCKER-FORWARD's one direct rule is scoped to -i docker0 (VLAN 50
arrives via enp1s0) - none can intercept VLAN 50 traffic regardless of
order. Fixed the script itself (delete + reinsert at position 1 if the
current first rule isn't ours) rather than loosening the Python
controller's forward_jump_first safety check, which was working
correctly when it refused to register against the drifted state.
Verified offline: rendered template passes bash -n, and the exact
live topology (rule present but 4th) reproduced against a bash mock
of iptables, confirming reposition-when-drifted, idempotency when
already first, and insert-when-absent all behave correctly. No real
iptables/wg/ip on this dev machine to test further; not yet
re-applied or re-verified against minisforum.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- Serialize every mutating command (recover, renew, register,
rewrite-config, pause, resume, state-reset) on the shared lock,
acquired BEFORE state is read rather than after: reading first and
locking only around the write let a slower command silently clobber
whatever a faster one wrote in between — in particular, an operator
pause could be lost to an in-flight recovery pass that read state
before the pause ran and only saved its own copy back afterwards.
- Fix emergency_close() to reposition its VLAN 50 DROP to the first
effective FORWARD position when it exists but has been pushed down
by something else (e.g. an out-of-band ufw reload) — a bare `-C`
existence check said nothing about position.
- Tighten wait_for_verified()'s handshake freshness check from >= to >
since_epoch: wg show and since_epoch are both whole seconds, so a
handshake at exactly the same second as since_epoch is
indistinguishable from a stale, pre-activation one and must not
count as proof of a new one.
- Fix the Ansible config-validation task to fail on any exit code
outside status's three documented health exits (0/1/2), not just on
5 — verified live against the real program (no wg/ip/iptables on
this dev machine reproduces the "nothing configured yet" first-run
case; status correctly returns 1, and broken/missing config.json
correctly returns 5 and is now caught).
- Correct the live fault-injection drill in the role README: port
1337 is only the addKey registration API port, not the tunnel's own
WireGuard UDP endpoint (derive that from `wg show pia-wg endpoints`
instead), and make the packet-capture check direction-aware
(`tcpdump -Q out`) so it can't conflate a real egress leak with
VLAN 50's normal inbound arrival on enp1s0.
Added 6 tests (86 total, all passing): emergency-close repositioning
past a preceding ACCEPT, pause/resume blocking on an in-flight
lock holder and surviving into a subsequent recovery pass, and the
handshake boundary at exactly/one-second-either-side of since_epoch.
yamllint, ansible-playbook --syntax-check, and ansible-lint (diffed
against main — no new finding categories) all pass. No live changes;
nothing pushed to main.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Enabling MSS clamping exposed a general hazard: systemctl restart
wg-quick@pia-wg tears down the live interface using whatever PreDown
rules are on disk right now, not what the interface was actually
brought up with. If Ansible had already rewritten pia-wg.conf with a
changed rule, the live interface (built under the old rules) won't
have whatever the new PreDown line tries to delete - iptables -D fails,
wg-quick's own set -e aborts the whole down/up sequence, and the
interface is left orphaned (never reaches its own built-in ip link
delete step), which then makes the following wg-quick up fail too with
"already exists".
Confirmed against wg-quick's real source (execute_hooks() runs each
hook via `(eval "$hook")`) that appending `2>/dev/null || true` to
each PreDown iptables -D line makes wg-quick treat it as succeeded
regardless of prior state - verified by rendering the template with
Jinja2 directly before trusting it against the live host. This
prevents the failure mode entirely for any future PostUp/PreDown
content change, not just this one.
Add INCIDENT-2026-08-25-vlan50-pia-qbittorrent.md covering all four
issues from this troubleshooting window: the zombie WireGuard tunnel,
a qBittorrent 5.2.0 stale-lock crash loop, the PMTU black hole this
fix addresses, and the restart failure hit while deploying it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
UFW allowed 53/udp+tcp from every other internal subnet but not from
10.10.0.0/24 (the WireGuard client subnet), so once the tunnel itself
was reachable, decrypted DNS queries still got dropped at minisforum's
own INPUT chain before reaching Technitium.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
AllowedIPs for the mac road-warrior peer was missing 10.10.40.0/24, so
DNS (10.10.40.53) and other home-LAN hosts were unreachable over the
tunnel. Also fixes the same string-vs-bool `when:` failure already
patched in e757850 (recent ansible-core rejects a `-e ...=true` CLI
string in a boolean conditional) for the client-config display tasks.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Creates enp1s0.50 on nik-debian (no L3 address, VLAN 40/node IP/default
route untouched) for Multus to later attach macvlan/ipvlan workloads to.
Verified live: enp1s0.50 is up with no IPv4 address (only the automatic
IPv6 link-local, which is expected and harmless), default route
unchanged.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Registers minisforum as a PIA WireGuard peer for VPN VLAN 50, with a
boot-ordered kill switch (dedicated PIA-VLAN50 iptables chain + a
terminal unreachable route in a dedicated routing table), multi-region
addKey fallback (Hong Kong -> Taiwan -> JP Tokyo, each region's full
server list, in order), and an observability-only health check.
Verified live against minisforum: registration succeeds, wg-quick@pia-wg
is up.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds explicit UFW routed-allow rules (10.42.0.0/16 pod-to-pod, pod-to-Technitium
DNS) since these nodes default their routed/FORWARD policy to DROP. Also brings
nik-debian's NFS (2049) and SMB (445) ports under Ansible-managed UFW rules via
the existing nfs-server role, scoped to the Lab/Trusted networks that need them.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
minisforum, debian, mac-mini, and gpu-node all moved from
192.168.7.0/24 to 10.10.40.0/24. Updates K3s server/agent config and
node IPs (including gpu-node's host_vars override), NFS export
allow-list and exports template, Pi-hole DNS records and
kube-vip/loadBalancerIP pins, WireGuard's pushed DNS/AllowedIPs, and
the NFS server IP baked into Jellyfin/Kavita/gitea-backup PVs and the
Ollama URL used by ai-gateway/Dashy.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This reverts commit fd13b8a. The registry-host migration broke image
pulls for home-services/portfolio (internal CA not trusted by
containerd) and has left them Degraded for 11 days; reverting back to
gitea.nik4nao.com pending a decision on how to proceed. Not pushed yet.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
The public domain is unreachable while moving, and the cluster had no
Traefik route to Gitea at all (public or internal), leaving every Argo
CD Application stuck in Unknown sync. Add a gitea.home.arpa
Certificate/IngressRoute, repoint Argo CD's repoURL, Gitea's own
DOMAIN/ROOT_URL/SSH_DOMAIN, the container registry references, the
Gitea Actions runner, and the watch-party clone URL at the internal
hostname.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
ansible.cfg sets inject_facts_as_vars: False repo-wide (the ollama role
already follows this via ansible_facts['system']), but these two tasks
referenced the bare ansible_hostname magic var, which doesn't exist under
that setting. Surfaced today as a hard failure re-running setup-gpu-node -
previously latent since these nodes' labels/taints were already correct
from whenever they last applied cleanly.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
host_vars/gpu-node.yaml's ufw_allowed_ports has overridden (not extended)
the common role's default list since the node was added, silently dropping
the Flannel VXLAN (8472/udp), K3s API (6443/tcp), and Kubelet (10250/tcp)
rules every other node gets. Went unnoticed because kubectl logs/exec/stats
tunnel through the agent's outbound connection to the k3s server rather
than needing a direct inbound path - but real pod dataplane traffic (e.g.
tts-gateway on nik-gpu resolving DNS against CoreDNS on nik-debian) needs
actual VXLAN connectivity and was blackholing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Deploys the Go gRPC gateway and Python/libtorch inference sidecar as one
pod on nik-gpu (nodeSelector/toleration/runtimeClassName: nvidia, sidecar
requesting nvidia.com/gpu: 1), matching the existing GPU device plugin.
Model checkpoint/config are bind-mounted from /data/tts-gateway, added to
gpu-node's Ansible data_dirs for consistency with the other hosts.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Stage 4 of REFACTOR_PLAN.md.
- Merge ansible/ansible.cfg into root ansible.cfg (single source of truth);
the dual-config setup silently broke documented commands and lost
host_key_checking=False when run from the ansible/ directory.
- Add ansible/requirements.yml pinning community.general, ansible.posix,
community.docker -- previously undocumented deps of the glances/watch-party
roles that would fail a fresh bootstrap.
- Align K3s version to v1.32.4+k3s1 across roles/k3s-server, roles/k3s-agent,
and host_vars/gpu-node.yaml defaults (was skewed: .2 vs .4). This only
changes what a *future* provision installs -- minisforum/debian are still
live on v1.32.2+k3s1 until separately upgraded.
- Fix kubeconfig fetch/replace path mismatch in k3s-server role: the `fetch`
task (flat: true) writes to ~/.kube/config, but `replace` was targeting a
/tmp/k3s-minisforum.yaml nothing creates -- would break a fresh rebuild.
- gitea-runner: only remove /run/docker.sock when it's actually a directory
(task name implied a check that wasn't there); tighten registration-token
systemd unit from 0644 to 0600.
- nvidia: stop unconditionally reporting `changed` (and restarting Docker)
on every run for an idempotent runtime-configure command.
- Gate the K3s join-token debug print and WireGuard client-config/QR display
behind opt-in vars (k3s_show_token, wireguard_show_client_configs), default
off -- both were printing real secrets to console on every run.
- Parameterize the docker role for Debian and Ubuntu; homeassistant now
depends on it (meta/main.yaml) instead of duplicating a Debian-only Docker
install inline.
- FQCN cleanup across wireguard, homeassistant, and ollama roles/handlers
(bare module names -> ansible.builtin.*/community.general.*/ansible.posix.*),
plus a few ansible-lint name-casing/idiom fixes. Handler renames verified
against their `notify:` call sites so notifications still fire.
- Update ansible/README.md and root README.md: add gpu-node/gpu_workstation
(4th host, previously undocumented), docker/nvidia roles, setup-gpu-node.yaml,
homeassistant.yaml, requirements.yml install step; correct the "Legacy"
homeassistant label (it's the only thing serving ha.home.arpa); correct the
Gitea ingress/backup-storage doc-drift; flag the Grafana/Loki static-PV
binding drift discovered via live cluster check.
Verified: all playbooks pass `ansible-playbook --syntax-check`, yamllint clean.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
feat(gitea): create PersistentVolume and PersistentVolumeClaim for Gitea
feat(gitea): add script to create Gitea runner registration token secret
feat(gitea): deploy Gitea Actions runner with Docker socket access
feat(media): deploy JDownloader with Ingress configuration
feat(media): set up Jellyfin media server with NFS and Ingress
feat(media): configure qBittorrent deployment with Ingress
feat(monitoring): add Grafana Loki datasource ConfigMap
feat(monitoring): create Grafana admin credentials secret
feat(monitoring): define PersistentVolumes for monitoring stack
feat(network): implement DDNS CronJob for Porkbun DNS updates
feat(network): create secret for Porkbun DDNS API credentials
feat(network): set up Glances service and Ingress for Debian node
fix(network): patch Pi-hole DNS services with external IPs
feat(network): configure Traefik dashboard Ingress with Authentik auth
feat(network): set up Watch Party service and Ingress for Mac Mini
refactor(values): update Helm values files for various services
- Created ansible.cfg for configuration settings
- Added inventory.yml for host definitions
- Implemented bootstrap playbook for Minisforum setup
- Developed setup playbook for K3s installation
- Defined common role with user and package management tasks
- Established K3s server role with configuration and installation tasks
- Included Traefik Helm values for ingress management