The internal + external monitor-sync CronJobs intermittently failed with socketio.exceptions.TimeoutError on api.login(), firing JobFailed -> Slack noise (and leaving monitor sync stale). Kuma 2.3.2 itself is healthy (1/1, 30m CPU); its single Node event loop just briefly stalls under ~300 monitors so the socket.io login handshake occasionally exceeds the client timeout. Wrap connect+login in a 5-attempt / 15s-backoff retry (disconnecting the half-open client between tries) so a transient stall no longer fails the whole job. Applied to both sync scripts.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Yesterday's Forgejo 3Gi->4Gi OOM fix pushed its tier-3-edge namespace quota (requests.memory=4Gi) to 100%, firing KubeQuotaAlmostFull + the healthcheck resourcequota check. Forgejo is the git + OCI-registry backbone and legitimately needs ~4Gi, so the edge tier's 4Gi ceiling is too tight. Opt the namespace out of the auto tier quota (resource-governance/custom-quota=true) and define a forgejo-specific ResourceQuota at requests.memory=8Gi, so the 4Gi pod sits at ~50% with headroom. Same opt-out pattern dbaas uses. Re-tiering was rejected: tier 1-cluster is also 4Gi, and 0-core (8Gi) would over-classify Forgejo's priority/eviction.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
k8s-portal was the last in-cluster image build; it now builds on GHA and
pushes ghcr.io/viktorbarzin/k8s-portal:latest, which is PRIVATE (infra repo
default). To pull it: add k8s-portal to the sync-ghcr-credentials Kyverno
allowlist (clones the ghcr-credentials Secret into the namespace) and
reference that secret via imagePullSecrets on the deployment — same wiring
as tripit/recruiter-responder. Completes the no-local-builds migration so
nothing builds container images on the cluster anymore (ADR-0002).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Recovered the real manifest + resolved lockfile (lockfileVersion 3, 71 pkgs)
from the running pod. A parent .gitignore force-ignored package.json, so the
git source tree was incomplete and the image only ever built manually. Now
reproducible on GHA (ADR-0002 no-local-builds).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
package-lock.json was never committed to either lineage — npm ci needs it,
so the build only ever worked from a manual devvm build with a local lock.
npm install resolves from package.json, unblocking the GHA build (ADR-0002).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The last in-cluster image build. GHA build-k8s-portal.yml builds
ghcr.io/viktorbarzin/k8s-portal:latest+sha (path-filtered on the Dockerfile
dir); Keel (force/poll/match-tag) rolls the deployment. Stack image repointed
to ghcr (ignore_changed); .woodpecker/k8s-portal.yml deleted.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Inline note on why the four backup CronJobs moved 10s->600s (bda1bdcb): a 10s deadline silently dropped the 2026-06-13 midnight full-backup run, firing PostgreSQLBackupStale. bda1bdcb rode in the same push as a forgejo change that failed CI on a namespace-quota error, so that pipeline failed before the dbaas apply took effect (live deadline was still 10s). This dbaas-only commit re-triggers the dbaas apply at a clean master so the 600s deadline actually goes live.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The 3Gi->6Gi bump in ff3cc44a was rejected by the forgejo namespace tier-quota (requests.memory capped at 4Gi). With Guaranteed QoS the 6Gi request exceeded quota; FailedCreate left forgejo with 0 pods for ~6 min (git remote + OCI registry outage) until I patched the live Deployment back to a schedulable 4Gi. 4Gi is the most the quota allows and is still a headroom bump over the OOM-prone 3Gi. To go higher the tier-quota must be raised in the same change. This reconciles TF to the live 4Gi so the pending/next apply is a no-op rather than reverting to the quota-busting 6Gi.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Forgejo OOMKilled twice on 2026-06-13 at the 3Gi cap (exit 137), briefly taking the git remote and OCI registry down and spiking ingress TTFB to 4.7s and the 4xx rate to 51%. Steady-state is ~2.2Gi but it spiked into the cap (true demand above 3.2Gi). The 2026-06-09 bump to 3Gi was sized for tripit buildkit registry pushes, but that driver is gone now that the Forgejo registry was frozen and emptied today (ADR-0002, images on ghcr), so the spike is git ops / the integrity-probe catalog walk / a possible leak. 6Gi gives headroom on the critical git backbone while we watch whether working-set keeps climbing (which would indicate a leak).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The daily full PostgreSQL backup silently skipped its 2026-06-13 00:00 run, leaving the last full dump 37h old and firing the critical PostgreSQLBackupStale alert. Root cause: startingDeadlineSeconds was 10s on all four dbaas backup CronJobs, so when the CronJob controller was more than 10s late to the midnight tick (many IO-heavy backups all fire at 00:00, the known etcd-starvation window) the run was dropped entirely instead of starting late. 600s lets a brief controller lag still launch the job. Applied to all four (mysql + pg, full + per-db) since they share the footgun and the midnight contention.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
ADR-0002 is fully landed (issues #11-#32 closed): every owned image now
builds on GitHub Actions and pushes to ghcr.io/viktorbarzin/<name>, with
Woodpecker reduced to deploy-only. The Forgejo container registry is frozen
and emptied; there are no in-cluster image builds or CI test runs anywhere.
The docs still described the old hybrid topology (DockerHub builds,
Woodpecker-native owned-app builds, the per-pattern migration lists, the
tripit-only pilot framing), which would mislead future sessions and
incident response.
This brings the docs to the completed reality (closes#33):
- docs/architecture/ci-cd.md: full rewrite as the canonical CI/CD reference —
the fleet GHA->ghcr->Woodpecker-deploy pattern, public/private ghcr package
split, infra-owned image workflows (incl. infra-ci on ghcr), the frozen
Forgejo registry, what Woodpecker still runs, and the #31 decommissions.
- .claude/CLAUDE.md: rewrite the "CI/CD Architecture" section to the
fleet-wide final state; FIX the stale claim that claude-memory-mcp builds
to DockerHub (it is GHA->ghcr); note owned images now live on ghcr and the
Forgejo registry is frozen/break-glass near the image-registry bullet.
- .claude/reference/service-catalog.md: f1-stream is GHA->ghcr + Woodpecker
deploy-only (was "Woodpecker-native build->deploy").
- stacks/{tuya-bridge,android-emulator}/variables.tf + stacks/terminal/main.tf:
cosmetic description/comment updates (forgejo -> ghcr; terminal-lobby has no
CI pipeline). Description/comment text only — no stack logic changed.
Historical records (docs/post-mortems/*, docs/plans/*) and ADR-0002 itself
are left untouched as point-in-time records.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
openclaw's install-nextcloud-todos-plugin init still pulled forgejo
nextcloud-todos (would ImagePullBackOff on restart once the forgejo
registry is wiped) -> ghcr:latest. f1-stream stack base (KEEL_IGNORE'd,
live already ghcr via set-image) repointed for fresh-create correctness.
Clears the last LIVE forgejo viktor/* refs before the registry reclaim.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
novnc's image was ignore_changed (KEEL_IGNORE) but nothing manages its
tag (keel.sh/policy=never), so the earlier forgejo->ghcr repoint never
took. Removing container[1].image from ignore_changes lets terragrunt
own novnc=ghcr:latest and roll it. container[0]/[2] (pinned playwright)
stay ignored.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Both now built by GHA → public ghcr. Repoint stack image bases
forgejo→ghcr:latest (terragrunt-managed, imagePullPolicy Always picks up
rebuilds). android var default api36-v8 -> latest.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Flips the planning workspace's Stay cover photos from the fake provider to live Wikipedia lead-image fetches (downloaded into STORAGE_DIR, served by the backend, editable per Stay). Part of the new-trip flow feature: every picked destination city gets a banner-ready cover. HOLD-ORDER: pushed only after the tripit image containing CityImageMode.wikipedia rolled out.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Service was already scaled 0/0 and unused (Viktor: 'not used anymore').
Live resources destroyed via scripts/tg destroy (10 resources: deployment,
namespace, service, anubis-travel + PDB/cm/svc/secret, ingress, TLS).
Removing the stack dir; old Woodpecker build (repo 5) deactivated
separately. The harmless legacy 'travel' CNAME->apex in config.tfvars is
left (now 404s; removing it would trigger a full-platform apply).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Formalizing x402-gateway CI (was a manual no-CI image). The deployment
lives in the traefik module; its image was NOT in ignore_changes, so a
set-image deploy would be reverted on the next traefik apply — added it
(KEEL_IGNORE_IMAGE). Base repointed to ghcr:latest; the GHA deploy
set-images the :sha8. Public ghcr package = no pull secret. Inert on the
live pod (image now ignored); rolling cutover keeps forwardAuth up.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
GHA now builds+pushes ghcr.io/viktorbarzin/claude-memory-mcp (public).
Image is KEEL_IGNORE_IMAGE (set-image managed), so this apply is inert
on the live pod; the stale :17 default literal is corrected to :latest.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
GHA now builds+pushes ghcr.io/viktorbarzin/claude-agent-service (public
package, anonymous pulls). Repointed: claude-agent-service (deployment +
git-init/seed-beads-agent inits), claude-breakglass, ci-pipeline-health,
beads-server CronJobs, k8s-version-upgrade (tag var 2fd7670d -> latest —
the Forgejo registry lost that sha; node caches were the only thing
keeping those CronJobs alive). publish-gate: vendor-contact emails
(licensing@/legal@/security@/sales@) ruled license-boilerplate, not PII.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
broker-sync is a CronJob-only consumer (no deployment): new --no-deploy
mode skips Woodpecker registration and renders build.yml without the
deploy job — :latest+Always CronJobs pick up builds on the next run.
wealthfolio stack: ghcr-credentials pull secret + image base repoint.
The wealthfolio-sync image regains a reproducible rebuild path.
Closes: code-62tm
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Second half of the recruiter-responder off-infra migration: the first GHA
build has published ghcr.io/viktorbarzin/recruiter-responder:{1d99a8d5,latest},
so the openclaw plugin-install init container can now follow the ghcr
:latest. The forgejo-side build pipeline was removed by the onboarding
commit, so the old forgejo :latest tag is frozen and would silently serve
stale plugin code. Deferred from the first commit on purpose - flipping it
before the package existed would have wedged the openclaw rollout on
ImagePullBackOff.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Migrating recruiter-responder off in-cluster Woodpecker builds: GHA will
build and push ghcr.io/viktorbarzin/recruiter-responder (PRIVATE package).
This commit lands the pull-side prerequisites BEFORE the first off-infra
build fires:
- stacks/recruiter-responder: image base forgejo -> ghcr (inert on the live
Deployment - both containers are ignore_changes'd; the Woodpecker deploy
moves the tag) + ghcr-credentials imagePullSecrets on the Deployment
(covers the recruiter-responder container AND the alembic-migrate init
container, which share the image).
- stacks/openclaw: ghcr-credentials imagePullSecrets on the openclaw
Deployment - its install-recruiter-plugin init container consumes the
:latest tag of this image. The image ref itself flips to ghcr in a
follow-up once the first GHA build has created the package (flipping now
would ImagePullBackOff on a not-yet-existing package and wedge the apply).
- stacks/kyverno: allowlist openclaw in sync-ghcr-credentials so the pull
secret is cloned into that namespace too.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Comment-only touch so the changed-stack detection applies
stacks/fire-planner from the current master tree. Pipeline 150 (commit
f18dfa4c — the ghcr image base + ghcr-credentials migration for issue
#26) was auto-killed when the concurrent nextcloud-todos push superseded
it, and pipeline 151 diffed from f18dfa4c onward so the fire-planner
stack changes were never applied (cronjobs still point at the forgejo
image, pod specs lack ghcr-credentials).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The nextcloud-todos build moved off-infra: GHA builds on the public
GitHub mirror and pushes ghcr.io/viktorbarzin/nextcloud-todos (public
package, anonymous pulls); Woodpecker repo 207 is deploy-only. First
ghcr image (:19c22d8c) is already built, deployed and rolled out, so
this repoint lands after the image exists. Both deployment image refs
(main + alembic-migrate init) are ignore_changes'd — no live churn,
the base matters only on resource (re)create. Old image was pulled
from a Forgejo registry package that no longer exists (pods survived
on node image cache only).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Migrating fire-planner off in-cluster Woodpecker builds to GitHub
Actions -> ghcr.io (ADR-0002, issue #26). The image base moves
forgejo.viktorbarzin.me/viktor/fire-planner ->
ghcr.io/viktorbarzin/fire-planner (a PRIVATE ghcr package), so the
deployment, all three cronjobs (recompute, col-refresh,
examples-weekly) and the examples bulk job gain the ghcr-credentials
imagePullSecret (the kyverno sync-ghcr-credentials allowlist already
covers the fire-planner namespace). registry-credentials stays
alongside so the currently-running sha-pinned forgejo image can still
be pulled until the first ghcr deploy lands; the cronjob images are TF
literals and flip to ghcr :latest on this apply.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Comment-only touch of both stacks so the changed-stack detection applies
them from the current master tree. Two pipelines went wrong in sequence
during the parallel ADR-0002 wave-2 migrations (issues #23/#24):
- pipeline 146 (instagram-poster stack prep, commit 29c69250) was
auto-killed when the concurrent payslip-ingest push superseded it, so
its apply never ran;
- restarting it as pipeline 148 inherited CI_PREV_COMMIT_SHA = the NEW
branch head (6928ce0b) with the OLD checkout (29c69250) — a reverse
diff that re-applied stacks/payslip-ingest from the pre-migration
tree, stripping the ghcr image base + ghcr-credentials pull secrets
that pipeline 147 had just applied (2 resources reverted).
This commit restores the committed payslip-ingest config exactly as
issue #24 landed it and finally applies the instagram-poster ghcr prep
from issue #23. Lesson encoded in the comments: do not restart killed
infra pipelines after master has moved — re-trigger with a touch commit
instead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Prep for moving payslip-ingest's image build off-infra to GitHub Actions ->
ghcr.io (ADR-0002 wave 2, issue #24). One stack commit before onboarding:
- image base repointed forgejo.viktorbarzin.me/viktor/payslip-ingest ->
ghcr.io/viktorbarzin/payslip-ingest (private ghcr package)
- ghcr-credentials imagePullSecrets added on the Deployment AND the
actualbudget-payroll-sync CronJob pod specs (namespace is already in the
kyverno sync-ghcr-credentials allowlist; secret verified present)
- the CronJob's SHA pin is retired: terragrunt image_tag 4f70681d -> latest
plus explicit imagePullPolicy Always on the cron container, per the fleet
convention for owned-app CronJobs — one less set-image target, and the
cron can never go back to pulling the dead Forgejo tag
The Deployment keeps KEEL_IGNORE_IMAGE; its concrete :sha8 tag is set by
the Woodpecker deploy pipeline after each GHA build.
Closes: nothing yet — the repo-side onboarding (offinfra-onboard) follows.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Prep for migrating instagram-poster off in-cluster Woodpecker builds to
GitHub Actions -> ghcr.io (ADR-0002, issue #23, PRIVATE-repo path).
Viktor asked for the wave-2 migration of instagram-poster per the wave-1
retro recipe: before onboarding, the stack must (a) carry the
ghcr-credentials imagePullSecret on the Deployment so the cluster can
pull the private ghcr image, and (b) repoint the image base from
forgejo.viktorbarzin.me/viktor to ghcr.io/viktorbarzin.
The Deployment image is KEEL_IGNORE_IMAGE (ignore_changes), so this
apply does NOT roll the pod to a not-yet-existing ghcr image — the live
forgejo-built :da5b4191 keeps running until the first GHA build POSTs
the Woodpecker deploy. The three CronJobs run curlimages/curl (public
DockerHub), not the app image, so they need neither the pull secret nor
a repoint. registry-credentials stays for the transition window.
Closes: nothing (stack prep only; repo onboarding follows)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
infra#17: the gate flagged npm deprecation boilerplate (package-lock.json
escapes the *.lock filter) and the upstream fork author's email in tracked
.beads data — both already-public upstream content, ruled false positives.
Lock files excluded properly; .beads moved to the eyeball inventory.
beads-server stack: beadboard image base repointed (deployment image is
KEEL-ignored; no CronJobs use it).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Freedify builds moved off-infra per issue #22: GitHub Actions on the
ViktorBarzin/freedify mirror now builds and pushes the public image
ghcr.io/viktorbarzin/freedify, and the Woodpecker deploy pipeline
(repo 202) rolls :sha8 via kubectl set image. Both factory deployments
(music-viktor, music-emo) now seed from ghcr instead of the retired
in-cluster Forgejo build, and the container image joins lifecycle
ignore_changes (KEEL_IGNORE_IMAGE) so terraform applies do not revert
the deployed :sha8. Landed after the first GHA push so ghcr :latest
already existed when this repoint applied. Public package - no pull
secret needed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
kms-website moves off in-cluster Woodpecker builds to GHA -> ghcr.
The kms-web-page deployment image is ignore_changes'd (CI sets the live
tag), so this repoint only governs future creates; package is PUBLIC so
no pull secret is wired. No CronJobs in this stack.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The legacy /settings/billing/actions endpoint now returns 410; sum
Minutes usageItems from /settings/billing/usage instead (found during
the infra#16 retro: June-to-date = 420/2000).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
publish-gate: gitleaks + trufflehog (full history) + PII heuristics;
CLEAN verdict gates any public flip, DIRTY = stays private. tuya-bridge:
ghcr-credentials pull secret + image base -> ghcr; namespace added to
the ghcr-credentials allowlist as a safety net (new ghcr packages
default PRIVATE even from public repos — prune after visibility flip).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
CronJobs track :latest via the TF literal (unlike the ignore_changes'd
deployment), so they kept pulling the dead Forgejo image after the
GHA/ghcr cutover — repoint the stack's image base.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
ADR-0002 wave 1 (infra#14): job-hunter's image moves to private ghcr;
the deployment AND both :latest CronJobs need the Kyverno-cloned pull
secret.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Stand up the infra for Viktor's break-glass: when the devvm is wedged (cluster
healthy), open breakglass.viktorbarzin.me, have Claude SSH in to diagnose/fix,
and power-cycle VM 102 via the Proxmox host if needed. App half landed in the
claude-agent-service repo.
New stack stacks/claude-breakglass/ — own namespace + SA, NO Vault role (ESO
syncs only its key, so the pod has zero direct Vault access). Hardened to
survive the pressure it exists to fix: priorityClassName tier-0-core, broad
node-pressure tolerations, anti-affinity off node1, imagePullPolicy Always.
auth="required" ingress so it rides the Authentik resilience proxy and stays
reachable via the basic-auth fallback during an auth-stack outage. Runs the
shared claude-agent-service image with the breakglass entrypoint.
files/breakglass-pve is the PVE forced-command (status|forensics|reset|stop|
start|cycle on VM 102, forensics-first).
Isolation: the shared claude-agent pod's terraform-state Vault policy is
explicitly DENIED secret/claude-breakglass/* (stacks/vault/main.tf) so a
prompt-injected agent on that pod can't read the root-on-devvm key.
traefik: add a checksum/auth-proxy-htpasswd annotation so the auth-proxy rolls
when the emergency basic-auth password rotates (it's a subPath mount that
doesn't auto-update) — regenerated this session so Viktor has a known
emergency credential, which the auth-stack-outage failure domain requires.
Docs: docs/runbooks/breakglass-ui.md (full incident + bootstrap procedure,
incl. the per-host from= NAT quirks) and a security.md note recording the two
new privileged footholds.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The 81a816f7 image (hybrid auth + OTA endpoints) is rolled out, so the
env can flip: AUTH_MODE=hybrid with the tripit-app OIDC knobs makes the
bearer-only tripit-api host actually authenticate Shell logins (browser
cookie path unchanged); BUNDLE_PUBLIC_BASE pins the signed OTA zip URLs
to that host; BUNDLE_TOKEN_SECRET joins the tripit-secrets ES (value
already written to Vault secret/tripit). Part of the Android APK work
(tripit #50/#51).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The sha tag other claude-agent-service CronJobs pin no longer exists in
the Forgejo registry (node caches mask it); fresh pulls 404. Follow the
owned-app CronJob convention until infra#19 moves this image to ghcr.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Review of tripit slice #50 caught that the provider's default
sub_mode (hashed_user_id) would make Shell JWTs carry a sub that
never matches the email-keyed prod user rows - first app login
would either 500 in placeholder reconciliation or split the user's
identity. sub_mode = user_email makes bearer and forward-auth
resolve the same row. Part of the Android APK work (tripit #50).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The devnen server runs chunked synthesis as a blocking call inside its
async handler, so the event loop (and every HTTP probe) hangs for the
whole multi-minute story. Kubelet's http liveness probe (1s timeout)
then killed the container mid-story (exit 137, twice within 10 min of
the first real drain), which reset the engine, so every following pass
started cold and tripit's 120s synthesis budget could never be met —
the queue would never drain.
TCP probes keep the meaning that matters: uvicorn binds 8004 only
after the model finishes loading in the lifespan hook, so readiness
still gates 'model loaded', while a GPU-busy server is left alive.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
HEAD~1 on a merge commit is the feature-branch parent, so the
changed-stack detection diffed the WRONG side and silently skipped the
stacks the push actually changed — pipeline 128 'succeeded' without
applying the new ci-pipeline-health stack. Use the push's true
before-state (CI_PREV_COMMIT_SHA) when it resolves, HEAD~1 as fallback
(first build / shallow edge cases). Also touches the ci-pipeline-health
stack so THIS push applies it.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Viktor asked to monitor the pipelines closely as builds move off-infra
(PRD infra#10). New aux stack: daily 07:30 UTC CronJob on the
claude-agent-service image running a deterministic shell sweep —
GitHub Actions failures/stuck runs across owned repos, Woodpecker
pipeline failures, GHA free-tier minutes burn. Healthy = one quiet
Slack line; issues = Slack alert + comment on infra#10. In-cluster
(not a cloud routine) because Vault + the Woodpecker token are
LAN-only. Secrets via ExternalSecret (github_pat deliberately, not the
ghcr_pull_token alias — a scoped packages-only rotation couldn't read
Actions runs).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
