The upstream manifest includes a Secret with empty client_id/client_secret
placeholders. We manage this via ExternalSecret, so drop the upstream copy
to avoid ownership conflicts in ArgoCD.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Use HTTPS raw URL from forge mirror instead of a separate ArgoCD app.
Pins operator image to v1.94.2 via kustomize images transformer,
avoiding the upstream's floating "stable" tag.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace vendored operator.yaml (495 KB) with ArgoCD apps sourcing the
upstream static manifest from mirrors/tailscale on forge, pinned to
v1.94.2 via targetRevision. Adds apps for both indri and ringtail
clusters. Local kustomization retains only ProxyClass and DNSConfig.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The 27B Q4_K_M model needs ~7.3 GiB system RAM for CPU-offloaded layers
but only 6.8 GiB was available within the 22Gi cgroup. Bumping to 24Gi
and enabling flash attention (reduces KV cache memory) should provide
enough headroom.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The 27B Q4_K_M model is ~17 GB, exceeding the 16 GB VRAM on the RTX 4080
by ~1 GB. Ollama will offload a few layers to CPU RAM, so the pod memory
limit needs headroom beyond the previous 16Gi.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Mosquitto has been dormant since frigate-notify switched from MQTT to
webapi polling (529ba10). Tear down live infra (ArgoCD app, namespace)
and remove all manifests, service-versions entry, services-check, and
doc references.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Add jobsync pod check (ringtail k3s) and HTTP endpoint to `services-check`
- Add JobSync entry to homepage dashboard under new "Apps" group
- Mark jobsync as reviewed at v1.1.4 (current with upstream)
- Changelog fragment added
## Deployment and Testing
- [ ] Sync homepage app from branch: `argocd app set homepage --revision review/jobsync && argocd app sync homepage`
- [ ] Verify JobSync appears on go.ops.eblu.me dashboard
- [ ] Run `mise run services-check` to verify new checks pass
- [ ] After merge: `argocd app set homepage --revision main && argocd app sync homepage`
Reviewed-on: #291
## Summary
- Add JobSync service reference card (`docs/reference/services/jobsync.md`) with architecture, secrets, observability, and JSearch API docs
- Add JobSync and Ollama to ringtail's workloads table (both were missing)
- Add JobSync to the reference index
- Wire `RAPIDAPI_KEY` through ExternalSecret and deployment env var for JSearch job search automation
- Document Loki log queries for observability (no metrics endpoint exists)
- Update deploy-jobsync how-to with new env var, observability section, and reference card link
## Deployment and Testing
- [ ] Sign up for RapidAPI JSearch API (free tier: 500 req/month)
- [ ] Add `rapidapi_key` field to "JobSync" 1Password item
- [ ] Merge PR
- [ ] `argocd app sync jobsync` to pick up new env var
- [ ] Verify job search works at https://jobsync.ops.eblu.me/dashboard/automations
Reviewed-on: #289
## Summary
C2 Mikado chain to deploy [JobSync](https://github.com/Gsync/jobsync) — a self-hosted job application tracker — to ringtail's k3s cluster.
### Mikado Graph
```
deploy-jobsync (goal)
├── build-jobsync-container
│ └── mirror-jobsync
└── integrate-jobsync-ollama
```
### What is JobSync?
Next.js app with SQLite for tracking job applications. Features resume management, application pipeline tracking, and AI-powered resume review/job matching.
### Key Decisions
- **Ringtail k3s** (not minikube-indri) — colocates with Ollama for zero-latency AI
- **Nix container** via `buildLayeredImage` — no Dockerfile, mirrors upstream source on forge
- **Ollama for AI** — uses existing deployment, no API keys needed for AI features
- **No upstream fork** — vanilla JobSync, Anthropic AI deferred to future work if needed
### Current Status
Planning phase — cards committed, ready for review before implementation begins.
Reviewed-on: #288
## Summary
- 1Password Connect uses non-standard numeric log levels (`1`=error, `2`=warn, `3`=info, `4`=debug, `5`=trace) per [1Password/connect#44](https://github.com/1Password/connect/issues/44)
- Alloy extracts the `level` JSON field as-is, so info-level health checks get `level="3"` in Loki
- Grafana expects string level labels — numeric values are unrecognized, causing misclassified log severity/coloring
- Adds a `stage.match` + `stage.template` in the Alloy pipeline scoped to `{namespace="1password"}` to normalize numeric levels to standard strings
- Other services are completely unaffected (scoped by namespace, not global)
## Deployment and Testing
- [ ] Sync alloy-k8s from branch: `argocd app set alloy-k8s --revision fix/onepassword-numeric-log-levels && argocd app sync alloy-k8s`
- [ ] Wait ~2 minutes for new logs to flow
- [ ] Verify level labels: `curl -sG "http://localhost:3100/loki/api/v1/label/level/values" --data-urlencode 'query={namespace="1password"}'` should show `"info"` and `"warn"` instead of `"3"` and `"2"`
- [ ] Check Grafana log panel for 1password namespace — logs should no longer appear as errors
- [ ] After merge: `argocd app set alloy-k8s --revision main && argocd app sync alloy-k8s`
Reviewed-on: #287
Frigate 0.17 does not auto-create clips/previews/<camera>/, causing
review page previews to silently fail with 500 errors.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Phase 2 of Dagger upgrade: bump engine version, update runner
deployment to v0.20.1-24f7512, and fix docs reference card version.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Bare image references in manifests were ambiguous — unclear whether the
tag was intentionally omitted or managed by kustomize. Add :kustomized
sentinel to all 37 image refs overridden by kustomize images transformer.
Add sync notes for tailscale-operator proxyclass (CRD fields not processed
by kustomize). Mark devpi reviewed (6.19.1 is current).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Built locally to break the chicken-and-egg: the old runner couldn't
build its own replacement because it needed Dagger 0.20.0.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The external-secrets operator adds conversionStrategy, decodingStrategy,
and metadataPolicy defaults to the live object, causing perpetual
OutOfSync in ArgoCD. Declare them explicitly to match.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Add Authentik OAuth2 provider + application blueprint for ArgoCD (ringtail side)
- Add OIDC config to ArgoCD ConfigMap with Authentik as identity provider (indri side)
- Map Authentik `admins` group to ArgoCD `role:admin` via RBAC policy
- ExternalSecrets on both sides pull `argocd-client-secret` from 1Password
- Local admin password remains as break-glass — both login methods coexist
## Pre-deployment manual step
Add `argocd-client-secret` field to "Authentik (blumeops)" in 1Password with a random value (e.g., `openssl rand -hex 32`).
## Deployment order
1. Sync Authentik app on ringtail first (blueprint + secret + worker env var)
2. Sync ArgoCD app on indri second (cm, rbac, ExternalSecret)
## Verification
- [ ] `argocd-client-secret` field added to 1Password
- [ ] Authentik app synced on ringtail — blueprint applied, provider created
- [ ] ArgoCD app synced on indri — OIDC config applied
- [ ] SSO login works: visit `https://argocd.ops.eblu.me` → "Log in via Authentik" → admin access
- [ ] Break-glass: local admin/password login still works
Reviewed-on: #284
Dashboard "Download/Upload Rate by Torrent" panels were querying
transmission_torrent_download_bytes (total_size * percent_done) and
transmission_torrent_upload_bytes (uploaded_ever) — cumulative byte
gauges, not rates. Added new metrics using Transmission's native
rate_download/rate_upload and updated dashboard queries.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The count-only stat wasn't actionable. New table shows pod name, container,
restart count, and memory limit for each OOMKilled container. Waiting reason
panel narrowed to make room.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
OOMKilled containers previously only appeared briefly in "Unhealthy Pods"
while dying, then vanished on restart. New panels use persistent metrics
(last_terminated_reason) and restart rate tracking.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Upgrade Transmission from 4.0.6-r4 to 4.1.1-r1
- Uses Alpine edge community repo for transmission packages, keeping stable alpine:3.22 base
- Fix stale image reference in service doc (was linuxserver, now custom registry image)
- Mark transmission as reviewed in service-versions.yaml
## Context
Service review found Transmission two minor versions behind (4.0.6 → 4.1.1). Alpine 3.22 only packages 4.0.6, so transmission is installed from edge's community repo with an exact version pin.
4.1.0 added improved µTP performance, IPv6/dual-stack UDP tracker, JSON-RPC 2.0 API. 4.1.1 is a bugfix release (20+ fixes).
Dagger test build passed locally.
## Deployment and Testing
- [ ] Build container via Forgejo workflow (`mise run container-build-and-release transmission`)
- [ ] Update kustomization.yaml with new image tag
- [ ] `argocd app set torrent --revision feature/transmission-review && argocd app sync torrent`
- [ ] Verify web UI at https://torrent.ops.eblu.me
- [ ] Check Grafana Transmission dashboard still receives metrics
- [ ] After merge: `argocd app set torrent --revision main && argocd app sync torrent`
## Note
The transmission-exporter sidecar (OOMKilling every ~30min, 294 restarts) is being tracked separately as a future replacement project.
Reviewed-on: #282
The database was at /config/frigate.db (emptyDir, ephemeral) instead of
/db/frigate.db (PVC, persistent). Every pod restart wiped the database,
losing all recording history and leaving orphaned files on NFS.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
ONNX detector + CUDA ffmpeg + workers consume ~1.9Gi at steady state,
causing intermittent OOMKills at the 2Gi limit.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Add two-stage Dockerfile for Loki (Go build → Alpine runtime) in `containers/loki/`
- Rewrite kustomize image to `registry.ops.eblu.me/blumeops/loki`
- Tag is `v3.6.5-placeholder` until first CI build; will be updated post-build
## Details
- UID 10001 matches existing StatefulSet `securityContext` (runAsUser/fsGroup)
- CGO_ENABLED=0, ldflags embed version via `github.com/grafana/loki/v3/pkg/util/build`
- Clones from `forge.ops.eblu.me/mirrors/loki` (mirror created this session)
- Pattern follows miniflux (two-stage Go) + prometheus (ldflags)
## Deployment and Testing
- [ ] Trigger container build: `mise run container-build-and-release loki`
- [ ] Update kustomize tag to actual build tag
- [ ] Deploy from branch: `argocd app set loki --revision feature/loki-container && argocd app sync loki`
- [ ] Verify `/ready` endpoint and log ingestion
- [ ] After merge: update to `[main]` tag (C0 follow-up)
Reviewed-on: #280
Was the only app still using https://forge.eblu.me (public proxy) for
git polling. All other apps already use the internal SSH endpoint at
forge.ops.eblu.me.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The k8s and ringtail runners were hitting forge.eblu.me (fly.io proxy)
for every FetchTask poll (~every 2s), round-tripping through the public
internet unnecessarily. Use forge.ops.eblu.me (Caddy on indri, tailnet)
for infrastructure workloads.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
