Changes for version 0.008 - 2026-09-22
- Add public `rest` and `new_object($kind, %fields)` on the client: the underlying `Kubernetes::REST` instance and a typed `IO::K8s` object for `create`/`update`, so callers no longer reach through the private `_rest`.
- Require `IO::K8s` 1.108 and `Kubernetes::REST` 1.108. Older IO::K8s resolved a bare Kind such as `Role`, `Event`, `Job` or `Service` to any loadable top-level package of that name -- `Class::Mite` ships a `lib/Role.pm` -- so creating the object died with `Can't locate object method "_k8s_attr_info"` on smokers carrying such a module (GitHub issue #2). The pair also brings UTF-8 request bodies (a manifest with a non-ASCII character died with "HTTP::Message content must be bytes"), the right API group paths for Storagemigration and Apiserverinternal, working inflation of KubeAggregator's `APIService`, and the upstream Kubernetes v1.37 sync with 11 new Kinds, all reached through the `Kubernetes::REST` seam with no change here.
- Accept a qualified `group/version/Kind` resource name anywhere the client takes one: `list('autoscaling/v1/HorizontalPodAutoscaler')` addresses that version, while the bare Kind keeps the shipped default of v2.
- Move `Net::Async::Kubernetes::PortForwardSession` into its own file. It was declared inline in `Net::Async::Kubernetes`, so `use`ing the class the POD of `port_forward()`, `exec()` and `attach()` names failed, and every `dzil build` warned about it. The code is unchanged.
- Add `on_watch_error` to `Net::Async::Kubernetes::Controller`. Watch `ERROR` events (a `403` mid-stream, for example) were wired nowhere; they carry a `Status` hashref rather than a keyed object, so they go to the callback instead of the reconcile workqueue. An `on_error` passed to `watch_resource` keeps precedence for that watch.
- Fix two reference cycles that leaked a controller and its client. The controller held a strong `kube` while the client held the controller as a child, and every reconcile context held a strong `controller` inside an entry the controller owns. Both are now weak, as in `Watcher`; a client the controller constructs itself stays owned by it.
- Fix the controller keeping one workqueue entry per key for the lifetime of the process, each pinning the last object it saw. A key is dropped once it reconciles cleanly; keys still queued, dirty or waiting on a retry keep their entry, so backoff history survives.
- Fix a controller key stalling after a stop/start cycle. `stop()` left the workqueue filled and its keys flagged queued or dirty, so the event a restarted watch delivered was deduplicated away with no drain scheduled, and a dirty key finishing its reconcile requeued past the stop. `stop()` now clears queue and flags together and the requeue skips while stopped; the failure counts stay, so a retrying key resumes at its next attempt.
- Fix `stop()` leaving a stopped watcher attached to the client on every stop/start cycle, so a repeatedly restarted controller piled up notifiers.
- Fix the duplex transport naming `port_forward` in its "not added to an IO::Async::Loop" failure whichever method was called; `exec()` and `attach()` now name themselves.
- Fix a resource name that `IO::K8s` cannot resolve killing the caller with "argument is not a module name" from `build_path`. `list()`, `get()`, `patch()`, `delete()`, `log()`, `port_forward()`, `exec()`, `attach()` and the controller's `patch_status()` now return a failed `Future` naming the resource; `$kube->expand_class` and starting a watcher croak with the same message.
- Stop `t/03-configure.t` reading the kubeconfig of the machine it runs on. One CPAN Testers smoker carried a kubeconfig with `contexts: null`, so every 0.005 report from it was a FAIL nobody else saw. The subtest now runs against a fixture in a temporary HOME, and a companion subtest holds the degenerate config to a catchable croak. Reported by Andreas Koenig, GitHub issue #1.
- Load `Net::Async::Kubernetes::Watcher` and `::Controller` at compile time instead of `require`ing them in `watcher()` and `controller()`: a broken child module now surfaces on `use Net::Async::Kubernetes`.
- Move the example scripts from `ex/` to `eg/`; the 0.007 entry's `ex/live_features.pl` ships as `eg/live_features.pl`.
- Documentation pass: PodWeaver generates the `NAME` sections, `PortForwardSession` is documented and in `SEE ALSO`, `list()` and the `cp` helpers describe what they do, the qualified resource name and the status-subresource write-strip trap are written down, and the `Controller` methods state the `Future` they return.
Modules
Async Kubernetes client for IO::Async
Minimal controller runtime for Net::Async::Kubernetes
Duplex websocket session for pod port-forward, exec and attach
Auto-reconnecting Kubernetes watch as IO::Async::Notifier