P0 Health-gated rollback (standalone auto-update path): capture the previous
image id + reference + healthcheck before the recreate, then poll the new
container's health over a configurable window. On healthy proceed (and only
then clean up the old image); on unhealthy/exit/timeout re-tag the old image
back onto the original reference and Recreate (no pull) to restore it, reusing
Recreate's config preservation. The decision is a pure decideRollback() helper.
P1 Per-endpoint enable: ContainerAutomationDisabled flag on Endpoint (zero value
participates, no migration churn), checked by both daemons; settable via the
endpoint update API. UI control deferred (see report).
P2 Notifier seam: minimal Notifier interface + logNotifier, emitting structured
updated/rollback/update-failed/heal-restarted events from the daemon.
Settings: RollbackOnFailure + RollbackTimeout (default 120s) added to
ContainerAutomation.AutoUpdate, wired through defaults/migration/golden,
settings_update validation, the AutoUpdatePanel and the TS types.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
F1: Stop routing git-backed stacks through a per-tick RedeployWhenChanged for
image-only updates. The git redeploy path short-circuits when the commit is
unchanged (so an upstream-digest update never applies) yet still git-fetches
every tick. Git stacks are now detect-only in the auto-apply path; their image
update lands on the next git change or via manual "Update now". File (non-git)
stacks still force-pull-redeploy immediately. The AutoUpdatePanel text no longer
promises daemon auto-update for git/externally-managed containers.
F2: Resolve registries for the file-stack redeploy the same way the established
userless/system path (RedeployWhenChanged) does, via the new
deployments.ResolveStackRegistries: scope to the stack author's endpoint access
and RefreshAndPersistECRTokens, instead of hand-passing Registry().ReadAll().
ECR-backed stacks now auto-update with fresh tokens.
F3: Add a 1m floor for the auto-update poll interval, enforced in the settings
Validate and mirrored in the frontend validation.
F4: Thread the application shutdownCtx into NewService and use it as the base
for the heal/update job operation contexts, so shutdown cancels in-flight work.
F5: Correct the updateEndpoint comment about monitor-only badge-cache warming
(only in-scope monitor-only containers are status-checked).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
claude code agent