[ferhimedamine]

Summary

Chart 0.12.3. Version and appVersion move 0.12.2 -> 0.12.3, and the dashboard tag moves 0.5.1 -> 0.5.2. The chart follows engine Dakera-AI/dakera#922 (#917 / #920 / #921 / #923) and includes the hardening found in a security audit.

Merge order: merge last.

1. Engine #922 first: it tags v0.12.3 and publishes ghcr.io/dakera-ai/dakera:0.12.3.
2. Then the clients and the dashboard 0.5.2 release (ghcr.io/dakera-ai/dakera-dashboard:0.5.2).
3. Then this PR and dakera-deploy.

Merging publishes chart 0.12.3, which needs both images to exist. helm-publish-oci.yml now skips versions that are already published, so 0.12.2 can no longer be overwritten.

Changes

Release

• Version bump with artifacthub images and changes, a CHANGELOG 0.12.3 section with upgrade notes, README version tables and install examples, "Upgrading from chart 0.12.2", and a NOTES "Server 0.12.3" block.

Server 0.12.3 settings (rendered only when set, range-checked in values.schema.json)

• dakera.config:
  • httpHeaderReadTimeoutSecs;
  • the recall switches (neighborhoodMode, recallRankDepth, ...);
  • maxConnections;
  • shutdownDrainSecs. The chart refuses it unless 15 + drain + 5 fits in the new dakera.terminationGracePeriodSeconds (30);
  • allowedHosts.
• config.documents.*: document import (DAKERA_DOCUMENTS, DAKERA_DOCUMENT_*).
  • PDF / DOCX extraction uses up to extractConcurrency x extractMemoryMb (2 x 1024 MB) inside the 4Gi limit. The docs say to raise the limit or use concurrency 1 for PDF-heavy imports.
  • ingress.annotations now defaults to proxy-body-size: 500m. ingress-nginx's own 1m default refused 25 MiB uploads.

Authentication off

• NOTES now warns that public dotted Host names (the ingress host) get 421 unless listed in allowedHosts, and prints a ready --set line. The Service FQDN (*.svc.cluster.local) is a private-use name the 0.12.3 server accepts, so in-cluster clients need nothing.

Behaviour of server 0.12.3 in the upgrade notes

• Autopilot dedup is off by default (DAKERA_AUTOPILOT_DEDUP_MODE=merge in dakera.extraEnv keeps it); a full queue answers 429 + Retry-After; recall of an agent with nothing stored answers 200 with no memories.
• Weight prepacking follows the memory limit: with the default 4Gi / 2 CPU the reranker's is on and the embedder's off (values.yaml, extraEnv).

Server image (distroless, uid/gid 1000)

• Probes stay httpGet and gain timeoutSeconds: 5.
• The GPU image is CUDA 13 and needs NVIDIA driver R580+ (documented).

Dashboard 0.5.2

• Runs as uid/gid 10001 with a read-only root filesystem, a /tmp emptyDir and no capabilities.
• New dashboard.trustedProxies (DAKERA_TRUSTED_PROXIES).

Hardening

• Read-only root filesystem by default; the data root is an emptyDir when persistence is off.
• RuntimeDefault seccomp and no service-account token on every pod.
• dakera.existingSecret.
• MinIO NetworkPolicy, on by default.
• Fixed the manual-Secret instructions.
• Actions pinned by commit SHA.

Tests

• scripts/validate-chart.sh covers all of the above, including 37 refusal checks. Its helm template | grep -q checks no longer fail with SIGPIPE under pipefail.

Testing

• scripts/validate-chart.sh (the CI job) passes with helm v3.14.4, checksum-verified.
• helm lint passes on the defaults and every example; helm package produces dakera-0.12.3.tgz.
• kubeconform -strict (k8s 1.30) finds 131 resources valid across the default render, all examples and a render with every option enabled.
• NOTES output checked for authentication off, with and without allowedHosts, and for upgrades.
• Against the 0.12.3 release commit's CI image (sha-c5b63fa): --check-config on the default render and on a render with every 0.12.3 setting set (plus throughput settings in extraEnv), also as a fresh install where an unknown DAKERA_* name is refused: every name is accepted. scripts/validate-chart.sh passes three runs in a row with helm v3.14.4.

🤖 Generated with Claude Code

https://claude.ai/code/session_019GcW157E6uJuFrBa2RgZBz