Three Junie briefs landed in one commit:
#07 — Wire cnpg-backup-sa into CNPG cluster spec (drift R8)
deploy/gcp/gke/knoe-db.yaml: add spec.serviceAccountName: cnpg-backup-sa
(requires CNPG v1.29+, which is the live operator version).
etc/init_cnpg_gke.sh: operator install URL now uses CNPG_OPERATOR_VERSION
variable (default 1.29.0); new §11 patches knoe-db and
knoe-db-barman-cloud RoleBindings to add cnpg-backup-sa as a subject
if not already present — matching the 2026-04-29 live stabilization.
#13 — Migrate off deprecated enablePodMonitor + podMonitorRelabelings
Both deprecated fields removed from deploy/gcp/gke/knoe-db.yaml
spec.monitoring. New deploy/gcp/gke/knoe-db-podmonitor.yaml carries
the PodMonitor with the cluster relabeling rule (cnpg.io/cluster pod
label → cluster label; required for all 85 CNPG Grafana panels).
Apply alongside knoe-db.yaml on next cluster patch.
#15 — Remove dead DASHBOARD consumer + basicauth_credentials
supabase/helm/knoe-supabase:
- wrapper.sh: drop DASHBOARD_USERNAME / DASHBOARD_PASSWORD envsubst lines
- config.yaml: drop DASHBOARD consumer + basicauth_credentials block
- kong/deployment.yaml: drop both DASHBOARD env-var secret refs
- values.yaml: rename secret.dashboard → secret.openai (apiKey only;
username/password dropped — no enforcer since commit 25f1b2e)
- secrets/dashboard.yaml + _helpers.tpl: renamed to openai /
supabase.secret.openai
- studio/deployment.yaml: reads from secret.openai.apiKey
- ci/example.yaml: updated to secret.openai.apiKey
helm template confirms knoe-supabase-openai secret referenced; no
DASHBOARD output.
docs/TODO.md: queue items #7, #13, #15 + drift row R8 archived to Done.
Co-authored-by: Junie <junie@jetbrains.com>
Defects A, B, C from docs/plans/junie/06-patch-garage-script-fixes.md:
A. DB_CLUSTER default was knoe-cnpg-0 (stale rebrand artifact);
corrected to knoe-dev-cnpg-0 to match conf/gke.cfg and the
cluster table in CLAUDE.md. Header comment + overrides block
also updated.
B. Phase 1 deletion loop missed service/garage-s3-ilb (the
LoadBalancer the ObjectStore endpoint historically pointed at).
Now included alongside service/garage.
C. Phase 2 was applying knoe-db-backup-gcs.yaml whole, including a
legacy ScheduledBackup using method:barmanObjectStore (being
removed in CNPG v1.30). Manifest split handled in prior commit
(34a25dd); script's Phase 2 is now safe to re-run.
bash -n clean; CONFIRM=false dry-run prints knoe-dev-cnpg-0 correctly.
docs/TODO.md: queue items #2 and #6 + drift rows R5 and R9 archived to Done.
Closes queue item #6 in docs/TODO.md (drift R9).
Co-authored-by: Junie <junie@jetbrains.com>
deploy/gcp/gke/knoe-db-backup-gcs.yaml previously contained both the
cnpg-backup-sa ServiceAccount (wanted) and a ScheduledBackup using
method: barmanObjectStore (unwanted — legacy in-tree method being
removed in CNPG v1.30, and the live cluster already has a plugin-method
ScheduledBackup that would be duplicated on re-apply).
Verified no other consumers of knoe-db-daily in the repo (grep clean).
File is now SA-only, matching its filename's intent.
Part 1/2 of queue item #6 fix (docs/plans/junie/06-patch-garage-script-fixes.md).
Co-authored-by: Junie <junie@jetbrains.com>
Mirrors commit b355855 (auth-side rename). The kustomization at
deploy/opentofu/k3s/manifests/knoe/kustomization.yaml has been
referencing knoe-* names since the rebrand; kustomize was broken
on the k3s deploy mode until this lands.
Five git mv:
prole-configmap.yaml -> knoe-configmap.yaml
prole-deployment.yaml -> knoe-deployment.yaml
prole-service.yaml -> knoe-service.yaml
prole-kdc-configmap.yaml -> knoe-kdc-configmap.yaml
prole-kdc-secrets.example.yaml -> knoe-kdc-secrets.example.yaml
Plus content sweep of knoe-kdc-configmap.yaml: replaced stale
PROLE.LOCAL/PROLE.ORG realm refs and PROLE_KDC_* env vars with
KNOE.DEV/KNOE_KDC_* to match the canonical GKE configmap.
Closes queue item #2 in docs/TODO.md (drift R5).
Co-authored-by: Junie <junie@jetbrains.com>
Bringing the long-running session-feature branch back into main in one
deliberate sweep. The branch carried the cluster work that's been live for
weeks (cross-cluster CNPG metrics, Grafana w/ Google OAuth, supabase
oauth2-proxy, cluster recovery, pg.0.knoe.dev + per-engineer onboarding,
GCS-backed CNPG backups via Workload Identity, the env-contamination
guard, the Junie brief queue, the cnpg-grafana CSRF + memory-request
fixes from today), while main accumulated Junie's parallel knoe-auth
Phase 2 OIDC work (full provider surface: discovery, authorize, token,
userinfo, JWKS, RS256 signing, code exchange, session services).
Key decision: the two branches did COMPETING rebrands off the same
starting point (5ba9b63, 2026-04-27):
- claude branch (commit b355855, earlier): org.prole.authority.* →
dev.knoe.auth.*
(artifact renamed to
knoe-auth.jar)
- main (commit 9daa94b, recent): org.prole.authority.* →
dev.knoe.authority.*
(kept "authority" artifact name)
dev.knoe.auth wins: cluster runs from this name, the Maven artifact is
already knoe-auth.jar, and the broader rename is the documented
namespace direction (per ~/.claude/projects/-Users-chrisfu-dev-knoe-db/
memory/MEMORY.md). All of main's recent Phase 2 OIDC content was ported
from authority/src/.../dev/knoe/authority/ into
authority/src/.../dev/knoe/auth/ with package declarations rewritten.
== File-level resolution summary ==
Textual conflicts (4):
authority/pom.xml
- Took our artifactId="auth"
- Took our branch's removal of spring-security-kerberos-client
(verified: Junie's Phase 2 OIDC code does not import it; the dep
was already-dead config)
docs/pipeline-phases.md
- Took our branch's "Phase 1 not started" status. Main had a
misplaced "✅ Complete" with a knoe-auth-Phase-1 commit ref
in the autobuild Phase 1 section — different domain.
docs/plans/knoe-auth-round-1.md
- Took our branch's dev.knoe.auth file table (vs main's
dev.knoe.authority listing). Pure rename mismatch.
supabase/helm/knoe-supabase/templates/kong/config.yaml
- Took our branch's onboard route + plain dashboard wiring.
Main had an oauth2proxy.enabled toggle that put oauth2-proxy as
a Kong upstream — but the deployed architecture (commit 25f1b2e)
has oauth2-proxy in FRONT of Kong, not behind. Main's wrapper
reflected an architecture that was never deployed.
- Took our branch's removal of basic-auth from dashboard route
(queue #15 brief still tracks the matching values.yaml /
kong/deployment.yaml cleanup).
Java tree reconciliation (44 file-pairs):
20 dual-path source files + 2 dual-path tests
Body-identical between main's authority/ and our branch's auth/
after stripping package decls — main's commit 9daa94b was a pure
rebrand. Took our branch's auth/ version for all 22.
8 main-only source files (Phase 2 OIDC), ported into auth/:
web/JwksController.java
web/OidcAuthorizeController.java
web/OidcDiscoveryController.java
web/OidcTokenController.java
web/OidcUserInfoController.java
session/OidcCodeService.java
session/OidcTokenService.java
session/SessionService.java
12 main-only test files, ported into auth/:
HealthControllerTest.java
enroll/EnrollValueTypesTest.java
enroll/EnrollmentControllerTest.java
enroll/TotpServiceTest.java
kerberos/KadminClientTest.java
kerberos/KerberosSpnegoResultTest.java
web/LoginControllerTest.java
admin/AdminControllerTest.java
user/PrincipalNormalizerTest.java
regression/IdentityRegressionTest.java
session/OidcCodeServiceTest.java
session/SessionServiceTest.java
Port mechanics: read main:authority/...<file> via git show, then sed
rewrite `package dev.knoe.authority` → `package dev.knoe.auth` and
`import dev.knoe.authority` → `import dev.knoe.auth`. Body content
unchanged.
authority/src/main/java/dev/knoe/authority/ — DELETED (duplicate)
authority/src/test/java/dev/knoe/authority/ — DELETED (duplicate)
== Verification ==
- grep -rln '<<<<<<<' across .java/.md/.yaml/.yml/.sh/.xml/.tpl: clean
- find authority/src -path '*/dev/knoe/authority*': empty (subtree gone)
- grep 'package dev.knoe.authority' across repo: clean
- bash -n install.sh deploy.sh etc/preflight_kubecontext.sh: clean
- git ls-files -u | wc -l: 0 unmerged paths
- helm lint supabase/helm/knoe-supabase: pre-existing failure on
studioIngress.enabled undefined in values.yaml (introduced by Junie
on main; unrelated to this merge — flagging as follow-up).
== Followups (carried into TODO ranked queue or noted here) ==
- helm lint failure: studioIngress block in values.yaml is missing
enable flag; templates/studio/{ingress,oauth2proxy-deployment,
oauth2proxy-service}.yaml all reference studioIngress.enabled with
no default. Pre-existing on main; not introduced by this merge.
- The five Junie briefs filed on this branch are now reachable from
main at docs/plans/junie/{02,06,07,13,15}-*.md. Junie can pick them
up in any order.
- knoe-auth Phase 2 OIDC source (now at dev.knoe.auth.*) is not yet
deployed to the cluster. Deployment is its own task.
- The branch claude/crazy-bose-fec256 stays in place (worktree at
.claude/worktrees/crazy-bose-fec256 may have ongoing context for
Claude Code sessions). Safe to delete once next session starts
cleanly from main.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Filed in response to the 2026-04-28 14:00 UTC backup outage. An
`install.sh --mode k3d` run with the shell pointed at GKE silently
overwrote the GKE cluster's GCS-backed ObjectStore + ScheduledBackup
with k3d-mode defaults; Garage filled up and CNPG backups failed for
hours before the next manual check. The class of bug is "config says
target cluster A, shell context says target cluster B, installer
proceeds against B without warning."
New shared bash helper at etc/preflight_kubecontext.sh with two
functions:
- verify_kubecontext_matches_config <cfg-path>
Strict gate. Reads [Global] APP_CLUSTER_KUBECONTEXT from the
config and exits 1 if `kubectl config current-context` differs.
Skipped silently when the config has no baked APP_CLUSTER_KUBECONTEXT
(e.g. fresh k3d.cfg) or when there's no live current-context.
- print_kubecontext_notice
Informational. Prints what's about to be inherited so the user
can abort before the TUI launches if it looks wrong. Never fails.
Wiring:
- deploy.sh sources the helper and calls the strict gate against
${PROLE_DEPLOY_CFG:-conf/gke.cfg} before invoking Python.
Unattended path -> hard refusal on mismatch.
- install.sh sources the helper and calls the informational notice
(gated on not-`--min`) right after entering the local-checkout
branch. The TUI is interactive, so the strict mode-aware gate is
a follow-up once the welcome screen records a mode in
state.inputs.
Bypass for deliberate cross-cluster maintenance:
KNOE_SKIP_KUBECONTEXT_GUARD=true ./deploy.sh
End-to-end verified:
- deploy.sh with current=cnpg-0, gke.cfg=app-0 -> exit 1, clear msg
- deploy.sh with KNOE_SKIP_...=true -> bypasses, prints
"skipping check"
- install.sh --min -> notice skipped
- install.sh (no flag) and install.sh --silent -> notice printed
Doc updates:
- CLAUDE.md §"Env-contamination warning" rewritten to describe the
live guard (was a forward-looking TODO).
- CLAUDE.md drift table row R4 removed; "Closed 2026-05-01" line added.
- docs/TODO.md queue item #1 archived to Done; R4 dropped from the
reality-vs-intent table. Queue numbering retained (no #1 placeholder)
so the docs/plans/junie/<NN>-...md filenames still match.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Five self-contained work briefs in a new docs/plans/junie/ subdirectory,
each tagged against a numbered item in docs/TODO.md so Junie can take
them independently in any order.
02-k3s-prole-rename.md -> queue #2 (drift R5)
06-patch-garage-script-fixes.md -> queue #6 (drift R9)
07-init-cnpg-gke-sa-wiring.md -> queue #7 (drift R8)
13-podmonitor-manual-management.md-> queue #13
15-remove-dead-dashboard-consumer.md-> queue #15
Each brief follows the same shape: Why -> What changes (concrete file
paths + line numbers + before/after) -> Verification -> Out of scope ->
Commit shape -> Definition of done. The intent is that Junie reads cold
(no shared chat history) and lands the change without escalating
questions.
Also adds:
- docs/plans/junie/README.md describing the convention.
- Row in docs/plans/README.md so newcomers find the subdirectory.
- Brief reference + "Assigned to Junie" tag on each of the five queue
items in docs/TODO.md.
Existing in-progress assignment to Junie (Phase 2 pg_oauth) is unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Implement Google OIDC support in Authority module via GoogleOAuthService
- Update AuthProperties and application.yml with OIDC configuration
- Add oidc-setup.md documentation for GKE/Google Cloud setup
- Update etc/init_knoe_auth.sh to handle OIDC secrets and path-B configuration
- Configure knoe-auth-deployment.yaml and gke.cfg for production auth
Co-authored-by: Junie <junie@jetbrains.com>
pytest_all.xml updated by IntelliJ to bind SDK name
'Python 3.14 (knoe-db)' and add PYTHONUNBUFFERED=1.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds authority module to root POM, updates Spring Boot and
dependency versions in authority/pom.xml, application.yml
updated for KNOE.DEV realm defaults.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Java implementation of the knoe-auth identity backbone (Round 1):
authority/.../admin/ user admin REST endpoints (list, deactivate,
reset-totp, role management)
authority/.../enroll/ invite-OTP web enrollment flow — accepts invite
token, creates Kerberos principal via kadmin,
seeds TOTP secret, returns keytab
authority/.../provisioning/ contributor provisioning service — orchestrates
invite generation, principal lifecycle, role grants
authority/.../kerberos/
KadminClient.java typed wrapper around the kadmin(1) subprocess;
create/delete/get-keytab/change-password
GKE manifests:
deploy/gcp/gke/knoe-auth-deployment.yaml two-container Pod (knoe-auth HTTP
+ kdc sidecar) with keytab-bootstrap
initContainer
deploy/gcp/gke/knoe-kdc-configmap.yaml krb5.conf, kdc.conf, kadm5.acl and
entrypoint for the embedded KDC
deploy/gcp/gke/knoe-kdc-secrets.yaml placeholder template — real values
created by init_knoe_auth.sh via 1Password
See docs/plans/knoe-auth-round-1.md for full architectural narrative.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds docs/plans/ as the canonical engineering reference for completed and
in-flight initiatives. Written for jr/mid engineers who have not seen the
repo before — each plan starts with strategic context and links to existing
code before asking for changes.
README.md index, audience, and status conventions
deployment-modes.md four-mode installer (min/k3d/k3s/gke), welcome-screen
mode selector, min-mode fast-path. Status: shipped.
knoe-auth-round-1.md Kerberos KNOE.DEV realm, invite-OTP enrollment,
Google corroboration, TOTP 2FA. Status: operational.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Delete tracked knoe-db/.prole_user_password (dead secret, rotated
in ansible since the original commit)
- Remove postgres-password.txt and build/ copy from working tree
- Add explicit ignore rules for *_password, postgres-password*,
.prole_user_password, *-secret variants
Add docs/plans/customer-deploy-resync.md with the upstream/downstream
migration plan and update CLAUDE.md to reflect prole's role as a
customer deploy of knoe-db.
Kong pod crashlooped on startup with:
init_by_lua error: plugin 'request-termination' not enabled;
add it to the 'plugins' configuration property
The chart sets KONG_PLUGINS as an explicit allowlist, and the
/healthz route we added in commit e61e6a8 uses the request-termination
plugin (bundled with Kong 2.8 but not enabled unless listed). Append
it. No image change, no new dependency -- just the plugin's name in
the comma-separated list.
This should be the last blocker: all the infrastructure for a healthy
supabase-kong backend is already in place (BackendConfig, checksum
annotation on the pod template, merged ingress with both hosts +
both ManagedCerts, the .invalid URL fix). With Kong able to parse its
declarative config, the pod can enter Ready, GCE's /healthz probe
returns 200, and api.0.knoe.dev / db.0.knoe.dev both start serving.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Second instance of the "chart Ingress ghosts out of the cluster shortly
after helm install" pattern. Earlier we worked around it for the kong
ingress by emitting a standalone manifest; studio/ingress.yaml now shows
the same symptom -- helm creates it, cluster reports `supabase-studio
is missing` seconds later, db.0.knoe.dev returns ERR_CONNECTION_CLOSED.
Rather than duplicate the externally-managed workaround for a second
ingress, this retires studio's ingress entirely and merges db.0.knoe.dev
onto the already-working supabase-kong ingress.
Kong's declarative config has always had a `dashboard` service block
that forwards `/` to supabase-studio:3000 (templates/kong/config.yaml).
So traffic for db.0.knoe.dev flowing through supabase-kong lands on
Studio exactly the same way it would have through the separate ingress,
just with one extra hop through Kong. From a user perspective:
identical. From our perspective: one ingress, one static IP, one
BackendConfig, no ghosting for Studio.
GCE supports multiple ManagedCertificates per ingress via a
comma-separated `networking.gke.io/managed-certificates` annotation;
each cert covers its own SAN. We already had both
supabase-api-managed-cert (api.0.knoe.dev) and
supabase-studio-managed-cert (db.0.knoe.dev) Active, so stitch them
both onto the shared kong ingress and SNI routes cleanly.
Changes in supabase/helm/render_supabase.py:
- `studioIngress.enabled` flips to false in k8s mode (chart's
studio/ingress.yaml is already gated on this value, so it no-ops)
- `service.studio.backendConfigName` cleared in k8s mode (chart's
studio/backendconfig.yaml and the studio Service annotation both
gate on this, so they no-op too -- single source of truth stays
in render_supabase.py)
- Standalone kong-ingress generator now merges studioIngress.hosts as
additional rules, same supabase-kong:8000 backend (Kong handles the
internal routing to Studio)
- Same generator merges studio's ManagedCertificate into the kong
ingress annotation as a comma-separated cert list
Studio's Service keeps existing (needed for Kong's in-cluster
`http://supabase-studio:3000` upstream). Only the public-facing Ingress
and BackendConfig go away.
Deployment-side follow-ups (not code):
- db.0.knoe.dev DNS must be repointed from 136.110.189.6
(the supabase-studio static IP) to 34.120.221.5 (the supabase-api
static IP, which is what this merged ingress is bound to).
- Once DNS is cut over, the supabase-studio global static IP is
orphaned and can be released to free the quota slot.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Follow-up to 23f3f14. With the checksum annotation in place, `helm
upgrade` did start rolling Kong forward, but the new pod never became
Ready: `rollout status` timed out with
"1 old replicas are pending termination"
Kong's ConfigMap contains the /healthz route correctly, the BackendConfig
is HTTP /healthz, but the Kong process itself refuses to start cleanly
when one of its services has `url: http://127.0.0.1:8000/` -- Kong's own
proxy port. This self-reference apparently trips the declarative-config
parser (undocumented; symptomatic only).
Switch to a RFC-2606 `.invalid` placeholder. The request-termination
plugin short-circuits before any DNS lookup or upstream connection, so
a non-resolvable URL is functionally equivalent and Kong's parser has
nothing to complain about.
Same change in both places we define the /healthz route:
- supabase/helm/knoe-supabase/templates/kong/config.yaml (supabase-kong)
- etc/init_kong.sh inline kong.yml heredoc (knoe-svc-kong)
The knoe-svc-kong backend was already HEALTHY with the self-referential
URL -- different Kong instance, possibly different parser path -- but
keeping the two configs aligned so future edits don't drift.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Follow-up to 0f2fe93. Live post-deploy diagnostics showed that even
though helm upgrade rewrote the supabase-kong ConfigMap with the new
/healthz route, the existing Kong pod kept serving the old config: the
Deployment spec itself never changed, so no rolling restart happened.
Kong loads /usr/local/kong/kong.yml at startup and doesn't watch the
file, so the healthcheck probe against /healthz got 404 and the GCE
backend stayed UNHEALTHY. api.0.knoe.dev stayed broken.
Standard helm workaround: annotate the pod template with a sha256 of
the ConfigMap template. Any content change bumps the hash, which
changes the Deployment spec, which triggers a rolling restart. Pattern
matches what vector/deployment.yaml in this chart already does and is
widely used in the bitnami / ingress-nginx charts.
Studio has no ConfigMap volume mount (verified via grep), so no
equivalent annotation is needed there.
This lands the permanent B.1 fix from the plan. Whether the live Kong
pod needs a manual `kubectl rollout restart` to pick up the current
REVISION 2 config -- or whether the /healthz service URL itself is
tripping Kong's parser -- is still to be determined from Part A
diagnostics before any B.2 / B.3 edits.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Follow-up to 391c4f5. Live deploy showed GCE's L7 BackendConfig CRD
hard-rejects type: TCP with:
Error syncing to GCP: error running backend syncing routine:
error ensuring health check:
Protocol "TCP" is not valid, must be one of [HTTP,HTTPS,HTTP2]
The sync never completes, so the LB has no healthy backend and TCP
connections to the public endpoint just close (ERR_CONNECTION_CLOSED).
Fix: switch all three BackendConfigs to type: HTTP with request paths
that return 200:
- supabase-kong & knoe-svc-kong: add a dedicated /healthz route to the
Kong declarative config via the request-termination plugin, which
returns 200 synchronously with no upstream call. Equivalent liveness
semantics to the TCP check we wanted (backend is alive as long as Kong
accepts connections) but over HTTP, which GCE actually accepts.
- supabase/helm/knoe-supabase/templates/kong/config.yaml
- etc/init_kong.sh (inline kong.yml heredoc)
- supabase-studio: Studio returns 301 on / (Next.js default) so we
point the probe at /favicon.ico -- Next.js serves it as a static asset
with 200 unconditionally. Not as clean as a real readiness endpoint
but Studio does not expose one that returns 200 without auth.
- supabase/helm/knoe-supabase/templates/studio/backendconfig.yaml
Verified locally via helm template -f values.generated.json: the
rendered BackendConfigs come out with the HTTP protocol + correct paths,
and the Kong ConfigMap has the healthz service block before the
auth-v1-open service.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
GCE L7 Ingresses for knoe-svc-kong (svc.knoe.dev / api.knoe.dev),
supabase-kong (api.0.knoe.dev) and supabase-studio (db.0.knoe.dev) were
all stuck UNHEALTHY: the default GCE healthCheck is HTTP GET `/` on the
backend port, but Kong returns 404 on any unrouted path and Studio
returns a 301 redirect -- neither passes the default probe, so the LB
serves "Server Error" instead of reaching the pod.
Replicate the pattern already working for gitlab-webservice-default:
emit a BackendConfig CRD with a TCP healthCheck on the service port and
annotate the Service with cloud.google.com/backend-config so GCE picks
it up. TCP is sufficient for LB-level liveness -- the backend is "alive"
as long as the process is accepting connections.
- etc/init_kong.sh: new SVC_KNOE_BACKEND_CONFIG_NAME; apply BackendConfig
inside k8s/GCE branch; annotate Service post-apply.
- knoe-supabase chart: new kong/backendconfig.yaml + studio/backendconfig.yaml
(TCP on 8000 / 3000), gated on service.{kong,studio}.backendConfigName.
- knoe-supabase chart: kong/service.yaml + studio/service.yaml pick up
cloud.google.com/backend-config when backendConfigName is set.
- render_supabase.py: sets service.{kong,studio}.backendConfigName in k8s
mode so the above wires up automatically.
Separately, the chart-managed supabase-kong Ingress was being reaped
from the cluster seconds after helm install (manifest present in the
release, gone via `kubectl get`). Root cause TBD -- suspected
meta.helm.sh/* annotation ownership colliding with a GKE/Anthos audit
controller. Workaround: render_supabase.py now emits a standalone
public-ingress-kong.yaml (no helm metadata) that supabase/deploy.sh
applies alongside public-ingress-tls.yaml, and the chart template gains
an `ingress.externallyManaged` guard so it no-ops in k8s mode. Default
`false` keeps k3d/k3s behavior unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The ansible excision in 5a769f2 was correct in scope -- no more host
inventory, no playbooks, no roles -- except it took out the one file
update.sh depends on for master DB password rotation:
infrastructure/inventory/group_vars/all/vault_db_master.yml
update.sh(1) reads this file (ansible-vault encrypted) at the start of
every run to decrypt the master password used to rotate k8s DB secrets
(knoe-db-user/superuser/knoe), ALTER ROLE on CNPG primary for the
prole/postgres/knoe/authenticator users, and reset Grafana admin creds.
With the vault file gone, update.sh failed with
ERROR: Vault file not found: .../vault_db_master.yml
and took down the tail of deploy.sh.
Restored from the parent commit of 5a769f2 verbatim -- content is
already ansible-vault AES256 encrypted, so it's safe to carry in a
private repo. The `.vault_pass` file that decrypts it remains
gitignored (never committed). Broader `infrastructure/` stays gone.
TODO (deferred): migrate this one remaining ansible-vault secret to
either OpenBao or a k8s Secret so knoe-db doesn't retain a residual
ansible-vault dependency surface. Tracked in our pending work.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Discovered today after the SSH-ingress reconcile: init_gitlab.sh had a
destructive false-positive in its "legacy storage detection" at the
top of setup_gitlab_storage(). The bare check
if kubectl -n "$NAMESPACE" get pvc repo-data-gitlab-gitaly-0 >/dev/null 2>&1;
then has_legacy=1; fi
flagged has_legacy=1 whenever the PVC object simply existed — which is
always true after the first successful deploy. Combined with
GITLAB_REPAIR_BLOCKED_AUTOCLEAN=1, every init_gitlab.sh invocation
triggered the "repairing legacy Gitaly storage" branch, which scales
the Gitaly StatefulSet to 0 and deletes repo-data-gitlab-gitaly-0
immediately, wiping all git-repo data.
We caught it when the knoe-db repo (pushed via HTTPS earlier in the
day) disappeared after an unrelated init_gitlab.sh run for SSH ingress
wiring. The PVC age on inspection was ~26 min — newer than the push.
git ls-remote against the project now returns a garbled gzip body,
consistent with an empty gitaly backing an existing DB project.
Fix:
- etc/init_gitlab.sh:setup_gitlab_storage()
* Remove the bare `pvc exists` legacy check.
* Keep the real legacy markers: synology PV, gandalf.prole.org
nodeSelector, gitlab-gitaly-static storageClass on the STS.
* Add a new check that flags has_legacy=1 only when the PVC's actual
storageClassName is a known-legacy value (gitlab-gitaly-static or
synology-iscsi), never on mere existence.
- conf/gke.cfg
* GITLAB_REPAIR_BLOCKED_AUTOCLEAN: 1 -> 0. The script fix above makes
re-enabling safe, but the conservative default is to let the user
see repair_blocked and consent rather than silently wipe data.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Course-correction on 82a2c2e: git.knoe.dev was already fully green at
34.107.159.254 (ingress Active, managed cert Active, DNS pointed at
it). Annotating the frontdoor ingress with
kubernetes.io/ingress.global-static-ip-name: git-knoe-dev would force
the GKE LB controller to swap the forwarding rule to the new reserved
address, which means a fresh IP, a DNS flip, and ~5-15 min of managed
cert re-Provisioning -- for zero functional benefit.
Blank GITLAB_GLOBAL_STATIC_IP_NAME in conf/gke.cfg. The surface
(env var handling in etc/init_gitlab.sh and the annotation conditional
in the CR render) stays intact -- it's just opt-out. When the gitlab
ingress next needs to recreate for another reason, reserve
git-knoe-dev with that IP before the recreate and set this key.
The three other ingresses (svc-knoe-ingress, supabase-kong,
supabase-studio) stay pinned, because they were all already in a
broken state (wrong DNS and/or Provisioning cert) -- swapping costs
nothing there.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Reserved 5 named external static IPs and threaded the
kubernetes.io/ingress.global-static-ip-name annotation through every
public GCE L7 ingress the installer renders, so ingress delete/recreate
stops churning DNS. Also turned on a dedicated regional L4
LoadBalancer for gitlab-shell (port 22) so SSH git workflows work.
New config surface in conf/gke.cfg:
GITLAB_GLOBAL_STATIC_IP_NAME = git-knoe-dev (34.102.141.87)
SVC_KNOE_GLOBAL_STATIC_IP_NAME = svc-knoe (34.111.197.33)
SUPABASE_API_GLOBAL_STATIC_IP_NAME = supabase-api (34.120.221.5)
SUPABASE_STUDIO_GLOBAL_STATIC_IP_NAME = supabase-studio (136.110.189.6)
GITLAB_SSH_HOST = git-ssh.knoe.dev
GITLAB_SHELL_LOADBALANCER_IP = 34.106.243.154 (regional us-west3)
GITLAB_SHELL_EXTERNAL_TRAFFIC_POLICY = Local (default)
Why SSH lives on a separate hostname+IP: Google-managed certs require
a GCE global L7 IP; port 22 needs a regional Network LB. Those can not
share an IP on GCP, so git.knoe.dev stays on HTTPS and git-ssh.knoe.dev
takes SSH.
Wiring per surface:
- etc/init_gitlab.sh
* Added GITLAB_SSH_HOST (default git-ssh.<root> in k8s mode),
GITLAB_SHELL_LOADBALANCER_IP, GITLAB_SHELL_EXTERNAL_TRAFFIC_POLICY,
GITLAB_GLOBAL_STATIC_IP_NAME config keys.
* CR global.hosts.ssh now reads ${GITLAB_SSH_HOST}.
* gitlab-shell block conditionally renders service: {type:
LoadBalancer, loadBalancerIP, externalTrafficPolicy} when the LB IP
is set.
* GITLAB_GCE_TLS_ANNOTATIONS_YAML gains
kubernetes.io/ingress.global-static-ip-name: "<name>" when
GITLAB_GLOBAL_STATIC_IP_NAME is set.
- etc/init_kong.sh
* Added SVC_KNOE_GLOBAL_STATIC_IP_NAME var near other SERVICE_TLS_*
defaults.
* gce_tls_annotations heredoc gets the static-IP line appended on the
same condition.
- supabase/helm/render_supabase.py
* Two new _first(env, cfg, default) extractions for
SUPABASE_API_GLOBAL_STATIC_IP_NAME and
SUPABASE_STUDIO_GLOBAL_STATIC_IP_NAME near the existing managed-cert
/ frontend-config vars.
* Two new dict-spread blocks in the Kong + Studio ingress annotations
that emit the static-IP annotation only when mode=k8s and the value
is non-empty.
All three ingress surfaces already use kubectl apply (merge-friendly);
re-running init_gitlab.sh / init_kong.sh / deploy.sh (supabase step) is
enough to pick up the new annotation. The GKE LB controller will swap
each ingress's forwarding rule from the auto-generated k8s2-fr-*
reservation to the named reservation, then release the old ephemeral.
Ingress IPs change; DNS records need updating (TTL <=300s while
iterating). Google-managed cert for gitlab-managed-cert will briefly
Provision again during the swap; the other three were already
Provisioning.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Follow-up to the ansible excision: remove other artifacts that should
never have been tracked, and extend .gitignore so they do not come back.
Removed:
- ssl/prole/*.key, *.crt, *.srl -- prole.org-era private keys, registry
and DB CA material. References in knoe/core/env.py (auto-detect path
at :1488) are best-effort inside try/except, so deletion is safe.
- ssl/prole.ssl.tar -- tarball of the same material.
- deploy/opentofu/k3s/terraform.tfstate -- 114KB committed OpenTofu
state snapshot with live cluster UIDs. tfstate is never supposed to
live in git.
- docs/deploy_gke.log, docs/deploy_gke_done.log,
docs/compressed-history.log -- deploy-time log captures.
- end_time.txt, start_time.txt, start_time_final.txt -- ansible wall-
clock timing breadcrumbs.
- network_description.txt, network_prompt.txt -- local scratch prompts.
- sidekiq_logs.txt -- transient log capture.
- test_resolve.sh -- ad-hoc debug script.
.gitignore additions:
- *.tfstate, *.tfstate.backup
- /ssl/
- .ai/
- /end_time.txt, /start_time*.txt, /sidekiq_logs.txt,
/network_description.txt, /network_prompt.txt, /test_resolve.sh,
/docs/deploy_*.log, /docs/compressed-history.log
Note: the purged SSL keys remain in git history until the deferred
history-scrub pass (task #14). The repo will be created as Private in
git.knoe.dev so this is acceptable for now; it MUST be scrubbed before
any broader visibility change.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
knoe-db deploys cluster-native (mode k3d for local dev, k8s for the
shared company cluster, k3s still supported for individual homelab
collaborators via conf/k3s.cfg). The ansible-managed distributed cluster
of prole.org machines is no longer the target topology, and the Ansible
inventory in this tree targets only the original author's specific hosts
(merlin/gandalf/myrddin/pi/raspberry/retropie.prole.org). It is neither
portable nor useful for the new deployment shape.
Removed:
- infrastructure/inventory/ (hosts.ini + 7 *.prole.org host_vars)
- infrastructure/playbooks/ (11 playbooks, all hardcoded to *.prole.org)
- infrastructure/roles/ (20 roles: pihole, samba AD DC, mariadb,
netplan, iscsi, prole_ssl, etc. -- all
tied to the bare-metal prole.org topology)
- infrastructure/setup/ (cr_ansible_user.sh, cr_samba_family_users.sh,
filesystem.txt -- prole.org bootstrap)
- infrastructure/deployments/svc-check-helm/
(legacy; etc/init_kong.sh already tears
down svc-check resources on every run)
- infrastructure/ansible.cfg (references deleted inventory)
- ansible.cfg, ansible.sh, ansible_min.cfg, ansible_recovery.log (root;
orphaned without inventory)
- vault_backup/ (ansible-vault encrypted prole.org secrets
backup: ad_dc_vault.yml, all_prole_vault.yml,
iscsi_vault.yml)
- tools/ (dashboard.sh, ghcr_docker_login.sh,
k3s-cluster-rename.sh, manage-node.sh,
export_mariadb_synology.sh,
prepare_mariadb_usb.sh, and a dumped
synology.prole.org mariadb .sql file --
all prole.org-era deployment helpers)
Kept: scripts/ (cluster ops), bin/ (prole-env.sh, prole-kpf.sh dev
shims), conf/k3s.cfg / conf/k3d.cfg (still supported modes; any
*.prole.org defaults in those files are legitimate mode fallbacks for
collaborators who keep running their own k3s/k3d setups). Test
fixtures referencing *.prole.org hostnames are left as-is (regression
data).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Enables iterating on GitLab reconciles without a full deploy.sh run and
unblocks the end-to-end login flow:
- etc/init_gitlab.sh
* Pre-scan $@ for --config before sourcing prole_cfg.sh so
PROLE_DEPLOY_CFG is set for standalone invocations; previously
prole_cfg.sh auto-picked conf/k3d.cfg and leaked KUBECONTEXT=dev,
failing with `error: context "dev" does not exist`.
* _init_gitlab_resolve_secretref: bash mirror of
knoe/core/actions.py:_resolve_secretref_value so
secretref://google-oidc-client-* resolves from etc/secrets/ files
when run outside deploy.sh Python env-injection.
* Garage bucket array: drop -storage suffix from uploads/artifacts/
lfs/packages/dependency-proxy. Chart defaults for these have no
suffix, and the object_store block in the CR does not override
per-object bucket names; the prior mismatch caused first-login 500s
(NoSuchBucket on avatar PUT).
- deploy/gcp/gke/gitlab-google-oidc-secret.example.yaml
* Add discovery: true so omniauth-openid_connect fetches Google
.well-known/openid-configuration; fixes "Could not authenticate
from OpenIDConnect: No host info" on the callback.
- conf/gke.cfg
* GITLAB_WEBSERVICE_LIMITS_MEMORY 1800M -> 3Gi, REQUESTS_MEMORY
900M -> 2Gi, REQUESTS_CPU 200m -> 500m. Live pod was sitting at
1706Mi/1800M (99%) in OOMKilled loop.
- conf/port-mapping.cfg: add supabase + gitea forwards, fix postgres
namespace knoe-db -> knoe-db-0 for split-cluster CNPG layout.
Adds read-only diagnostics used to chase the above:
- etc/diag_gitlab_boot.sh, diag_gitlab_webservice_oom.sh, diag_gke_storage.sh
- etc/ensure_default_storage_class.sh + k8s/prole/storageclass-gcp-standard-hdd.yaml
(preflight + HDD-default SC manifest for SSD-quota-constrained GKE projects).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
