Prevents OOM when switching between models — only one 14B model
fits in 16GB VRAM at a time with KV cache for context.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Replace pre-commit with [prek](https://github.com/j178/prek), a faster Rust-native drop-in alternative
- Migrate config from `.pre-commit-config.yaml` (YAML) to `prek.toml` (TOML)
- Add new built-in checks: case conflicts, private key detection, executable shebangs
- Install prek via mise native registry (`aqua:j178/prek`) instead of pipx
- Update all doc references across README, contributing guide, and how-to docs
## Notes
- `check-yaml` still uses the remote `pre-commit-hooks` repo because prek's builtin fast path doesn't support `--unsafe` yet (needed for Ansible custom YAML tags)
- All existing custom hooks (docs validation, container version check, mikado invariant, workflow validation) work unchanged
- Tested: all hooks pass on clean tree, deliberate doc link breakage is caught
## Test plan
- [x] `prek run --all-files` passes all checks
- [x] Broken wiki-link correctly caught by `docs-check-links`
- [x] taplo-format auto-fixes TOML formatting on commit
- [x] commit-msg hook (mikado invariant) fires correctly
Reviewed-on: https://forge.ops.eblu.me/eblume/blumeops/pulls/276
The mikado-branch-invariant-check hook now inspects staged files (commit-msg
hook) and historical commit files (standalone mode) to reject impl commits
that touch markdown files with Mikado frontmatter (requires:, status:, or
branch: mikado/). Cards should only be modified by plan, close, or finalize.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Fix go-server-derivation: wrong path target (webui not authentik-django)
and missing internal/web/static.go patch. Remove stale DRF fork content
from mirror-build-deps (no longer needed as of 2026.2.0). Add
last-reviewed to all 5 cards without it.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Python 3.14's EntryPoints uses string keys, not integer indices.
eps[0] raises KeyError(0); use next(iter(eps)) instead.
Verified on ringtail with the actual venv python.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Django's BASE_DIR is $out but source lives in site-packages. Code
like BASE_DIR / "authentik" / "sources" / "scim" / "schemas" / ...
needs a top-level symlink to find data files alongside Python source.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
lifecycle/ak uses ${TMPDIR}/authentik-mode — without TMPDIR set it
tries to write /authentik-mode in root, which user 65534 can't do.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
buildLayeredImage doesn't create /tmp by default. The container runs
as user 65534 (nobody) which can't mkdir /tmp at runtime.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Update image tag to v2026.2.0-efa9806-nix — the first source-built
authentik container from the build-authentik-from-source chain.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Create goal card and 4 prerequisite cards for building authentik from a
custom Nix derivation instead of using pkgs.authentik from nixpkgs. This
removes the dependency on the nixpkgs packaging timeline and gives full
version control over authentik releases.
Chain: mikado/authentik-source-build
Leaf nodes: authentik-api-client-generation, authentik-python-backend-derivation
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace floating :18 tag with pinned :18.3 (upstream out-of-cycle
release fixing 18.2 regressions). Stamps service as reviewed.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The LLM should read the file itself using its tools rather than
receiving it inline in the task output.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The external-secrets webhook injects conversionStrategy, decodingStrategy,
and metadataPolicy defaults on admission. Declaring them explicitly prevents
ArgoCD SSA from flagging the resource as OutOfSync.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Add `cluster` label (indri/ringtail) to all Prometheus scrape jobs, Alloy k8s metrics/logs, and Alloy host metrics/logs
- Deploy kube-state-metrics on ringtail's k3s cluster (ArgoCD app + manifests)
- Deploy Alloy on ringtail to collect pod metrics and logs, remote-writing to indri's Prometheus and Loki
- Replace single-cluster "Minikube Kubernetes" and "K8s Services Health" dashboards with:
- **Kubernetes Clusters** dashboard — multi-cluster with `cluster` and `namespace` template variables
- **Ringtail (k3s)** dashboard — dedicated ringtail view with GPU usage panels
## Deployment and Testing
1. Sync `apps` on indri ArgoCD to pick up new app definitions (`kube-state-metrics-ringtail`, `alloy-ringtail`)
2. Sync `prometheus` → verify `cluster` label on scraped metrics
3. Sync `alloy-k8s` → verify `cluster=indri` on remote-written metrics and logs
4. Run `mise run provision-indri -- --tags alloy` → verify `cluster=indri` on host Alloy metrics/logs
5. Sync `kube-state-metrics-ringtail` → verify pods running on ringtail
6. Sync `alloy-ringtail` → verify pods running, check Prometheus for `kube_pod_info{cluster="ringtail"}`
7. Sync `grafana-config` → verify dashboards appear, cluster variable populates both values
8. Check Loki for `{cluster="ringtail"}` logs from ringtail pods
## Notes
- Alloy on ringtail uses `insecure_skip_verify=true` for TLS to Prometheus/Loki (Tailscale-managed certs not in container trust store) — tighten later
- DNS resolution for `*.tail8d86e.ts.net` from ringtail pods depends on CoreDNS inheriting host's MagicDNS resolver; may need CoreDNS forwarding rules if pods can't resolve
- The old services dashboard (blackbox probes) is removed — those probes are still running in alloy-k8s and the data is still in Prometheus, just not in a dedicated dashboard
Reviewed-on: https://forge.ops.eblu.me/eblume/blumeops/pulls/270
Parked car was being re-detected every few minutes at night due to IR
illumination noise triggering motion detection. Restrict the driveway
zone to [person, dog, cat] so cars and birds no longer create events
there. Cars still alert via the driveway_entrance zone.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- **mirror-create**: Auto-includes GitHub PAT from 1Password for authenticated upstream fetches at mirror creation time
- **mirror-update-pats**: New mise task that SSHes into indri and rewrites the git remote URL in every GitHub mirror's bare repo config to embed the PAT. Idempotent, supports `--dry-run`
- **app.ini.j2**: Explicit `[mirror]` section with `DEFAULT_INTERVAL = 8h` and `MIN_INTERVAL = 10m` (bakes in the defaults for visibility)
- **manage-forgejo-mirrors**: New how-to doc covering mirror creation, PAT storage, the `mirror-update-pats` task, and the full 20-day PAT rotation procedure
## Context
GitHub tightened unauthenticated rate limits for git clone/fetch in May 2025. With 23 GitHub mirrors syncing every 8 hours, authenticated fetches avoid throttling. The PAT is stored in 1Password (`Forgejo Secrets` → `github-mirror-pat`) and has been applied to all existing mirrors.
## Deployment and Testing
- [x] `mirror-update-pats` dry-run verified (23 mirrors detected)
- [x] `mirror-update-pats` applied to all 23 GitHub mirrors on indri
- [x] Idempotency confirmed (re-run shows 0 updated, 23 skipped)
- [ ] Provision indri with `--tags forgejo` to apply `[mirror]` config
- [ ] Trigger a manual mirror sync and verify success in Forgejo UI
Reviewed-on: https://forge.ops.eblu.me/eblume/blumeops/pulls/269
The --style=header --color=never --decorations=always flags are now built
into the script so callers can just run `mise run ai-docs`. Also adds a
note to CLAUDE.md to never truncate the output.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Point ArgoCD app directly at forge-mirrored upstream repo (`mirrors/cloudnative-pg`) instead of the Helm charts repo
- Use `directory.include` to select the specific release manifest (`cnpg-1.27.1.yaml`) from the `releases/` directory
- No vendored files, no Helm — upgrades are a two-line change (`targetRevision` + `directory.include`)
- Delete unused `values.yaml` (was empty, all Helm defaults)
## Deployment and Testing
- [ ] Register mirror repo in ArgoCD: `argocd repo add ssh://forgejo@forge.ops.eblu.me:2222/mirrors/cloudnative-pg.git --ssh-private-key-path <key>`
- [ ] `argocd app set cloudnative-pg --revision feature/cnpg-direct-source && argocd app sync cloudnative-pg`
- [ ] Verify operator pod running: `kubectl get pods -n cnpg-system --context=minikube-indri`
- [ ] Verify CRDs exist: `kubectl get crd --context=minikube-indri | grep cnpg`
- [ ] Verify existing clusters healthy: `kubectl get clusters -A --context=minikube-indri`
- [ ] After merge: `argocd app set cloudnative-pg --revision main && argocd app sync cloudnative-pg`
## Notes
- The forge mirror was created via `mise run mirror-create` from `https://github.com/cloudnative-pg/cloudnative-pg.git`
- ArgoCD may need the mirror repo added to its known repositories if the credential template doesn't already match `mirrors/*`
Reviewed-on: https://forge.ops.eblu.me/eblume/blumeops/pulls/268
