v0.2.0

Install

helm install continuo oci://ghcr.io/carolsimone/charts/continuo \
  --version 0.2.0 -n continuo --create-namespace

Added

• remediation-agent.env.REMEDIATION_PR_OPENING_GRACE_PERIOD (default "10m") — how long the remediation-agent reconciler's opening sweep waits, measured against the claim's own pr_claimed_at, before releasing a stranded pr_state='opening' proposal back to 'failed' for retry. Sits alongside the existing REMEDIATION_PR_POLL_INTERVAL in the same free-form env map, so no schema change is required; an unmodified existing values file already gets this default via the chart's own env defaults.
• validation.imageTag — pins the external continuo-validation-<engine> image version independently of the chart appVersion (default v0.2.0). The key is optional (not in validation's required list), so a release that predates it — e.g. one upgraded with helm upgrade --reuse-values — still validates and falls back to the chart's current default in templates/_helpers.tpl's continuo.validation.image. An explicit imageTag: null override collapses to that same "key absent, use the default" path (Helm drops a null-valued override key before merging). A present-but-empty string ("") is different — it survives the merge as a deliberate misconfiguration — and is rejected: at helm lint/install time by the schema's minLength: 1, and at helm upgrade --reuse-values time (where the schema cannot see the merged value) by a fail in the same helper. Accepts an optional @sha256:<digest> suffix ("vX.Y.Z@sha256:<digest>") for an immutable pin — a plain tag is mutable, since continuo-validation's publish workflow re-pushes :vX.Y.Z on every tag; see values.yaml/README.
• helm upgrade NOTES.txt now reminds operators that the validation image is externally released and pinned by validation.imageTag independently of appVersion, and that private-registry mirrors need it mirrored explicitly — an unmirrored bump otherwise renders healthy at install and only fails later, at the next release promotion, as a validation-pod ErrImagePull.

Changed

• The validation image is now the externally released continuo-validation-<engine> (from github.com/carolsimone/continuo-validation), replacing the chart-appVersion-tagged continuo-validation-runner-<engine>; global.imageTag no longer applies to it.

Fixed

• Service pods now carry a checksum/config annotation over the shared
  ConfigMap, so helm upgrade restarts the pods whose configuration actually
  changed. Services read that ConfigMap through envFrom, and Kubernetes never
  refreshes environment variables in a running pod, so changing REDIS_HOST,
  S3_BUCKET, LOG_LEVEL or validation.engine previously rendered an
  identical Deployment and every pod kept serving the old value indefinitely.
  The engine case corrupted output rather than merely going stale:
  executor-controller rolls on an engine change because its validation image
  reference changes, so manifest-controller would have kept uploading the
  previous engine's SQL to the new engine's validator. Upgrade note: an
  upgrade that changes any shared ConfigMap value now performs a rolling
  restart of the affected services; an upgrade that changes none does not, as
  the digest covers only values derived deterministically from values.yaml
  (generated credentials live in Secrets).
• manifest-controller now reads and re-renders SQL in the dialect of the
  engine the install actually targets, instead of always assuming Postgres. The
  shared ConfigMap gained a WAREHOUSE_ENGINE key derived from the existing
  validation.engine value — no new values key, so unmodified overrides keep
  working. On a validation.engine: trino install, the candidate SQL uploaded
  for blue/green validation was previously re-rendered through sqlglot's
  postgres dialect, emitting constructs Trino rejects (a cast came out as
  CAST(x AS TEXT) rather than CAST(x AS VARCHAR)), so validation could fail
  on SQL the warehouse would otherwise have accepted. An engine with no dialect
  mapping now fails the service at startup rather than silently emitting another
  engine's SQL.
• Redis-backed services (manifest-controller, ui-service) no longer start
  before the bundled Redis is reachable. They previously raced the Redis
  StatefulSet, crashed on connect, and entered CrashLoopBackOff — whose
  exponential backoff (10s, 20s, 40s, 80s...) kept them down well after Redis
  became ready, so the restart loop cost more time than the wait it replaced. A
  wait-for-redis initContainer now gates them, mirroring the existing
  wait-for-migrations gate for Postgres-backed services. Bundled installs only;
  BYO installs point at Redis that is already running and render no initContainer.
• The quickstart's /etc/hosts step is documented with the two things operators
  actually get stuck on: it must be run from a real terminal, because sudo reads
  the password from the controlling terminal and exits with a terminal is required to read the password in an IDE pane or agent shell; and the password
  it asks for is the local account password, not the Dex demo login printed
  directly below it. The chart README now also documents a browser
  --host-resolver-rules flag that resolves the issuer hostname without root,
  for operators who cannot use sudo at all.