Where it was tested: the SUSE VPN is down, so the standalone previews load but can't reach a Harvester. Every live check ran on the PR's extension build loaded into mf's Rancher v2.16 in a test browser (the installed harvester extension blocked there only), on harvester-11-2 (.105; its API now reports master-9dd27fdb-head, a side-loaded master build on v1.10.0-rc1) and harvester-118 (.118, v1.9.0-rc4). Every save was held in the browser and every session logged out. Both previews are rebuilt at a3392e7 and f3ca9e4 for when the VPN is back.
Approvegrade A-
Both comments from your 11-10 review are fixed, proven with the same checks that found them, and the new commit adds no problem of its own. CI is green, the conflict is gone, and the PR is blocked only on an approval. Approving now is safe: no push is pending.
Screen reader names: Addressed. Both groups now carry their own name, Restore VM's too, as you asked.
Save test: Addressed. Both breakages that passed on 11-10 now fail the spec, and so do dropping or swapping either label.
Your question in the review body is still unanswered on GitHub:"Have we tested as an extension on rancher?". We have, in rounds three and four, so the approval's intro can answer it.
One UX point is your call (I'd leave it out): the form still doesn't say what either value does. It's unchanged since you reviewed it twice, and the author removed exactly that kind of warning on purpose in round two. Details under "Seen, not for the review".
The minus in the grade is that help-text gap, not anything in the new commits.
Screen readers get no name for this group: the bold label above it isn't tied to it, so it's announced as just "Enabled / Disabled". Restore VM right above has the same gap, so the two can't be told apart. Passing each label's text to its `RadioGroup` as `aria-label` would name both without changing how they look.
[image: accessible names of the two radio groups]
addressed in https://github.com/harvester/harvester-ui-extension/pull/1218/changes/a88dde33d7cbd3df4a0122f0ce0c985a07c01518
What changed
Both RadioGroups get :aria-label with their own label's translation key (lines 189 and 200). The reply links a88dde3, the rebased feature commit; the change is in a3392e7. Harmless.
Checked the same way as on 11-10
Chromium's accessibility tree of the live form in Rancher: main has one unnamed radiogroup (Restore VM); the PR build has radiogroup "Restore VM" and radiogroup "Shut Down VM on Migration Failure". A real Tab walk lands on each by name, Arrow Up picks Enabled, and on the v1.9 cluster Restore VM is named too. The name is the visible label's own text. axe 4.10.2: 0 violations. Both passes got the same result.
On 11-10 (d73d8aa): both groups unnamed. Round three's capture.
Now (a3392e7), in Rancher on harvester-11-2: both named.
Nit, optional: this test sets `parseDefaultValue` and calls `update()` itself, so it never goes through the radio group. Removing `@update:value` from the new group, or binding it to `restoreVM` by mistake, still passes all five tests. Emitting `update:value` from the group's stub would catch both.
[image: deliberate breakages run against this spec]
addressed in https://github.com/harvester/harvester-ui-extension/pull/1218/changes/a88dde33d7cbd3df4a0122f0ce0c985a07c01518
What changed
"should save the value selected in its radio group" now emits update:value from each group's stub and checks both fields; a new test checks both names. Spec 6/6 and feature-flags 7/7 pass; CI lint and unit tests are green.
Checked the same way as on 11-10
Breaking the component one way at a time and running the spec, at the old head and at the new one (Pass B's table). Both breakages from your comment pass the old spec and fail the new one; dropping or swapping either label and dropping Restore VM's handler now fail too. Pass A ran the same breakages on the new spec and got the same result.
Pass B: each breakage against the spec at d73d8aa and at a3392e7.
See it
The change is invisible on screen, so the clip Tabs through the form for real and puts in the caption what the accessibility tree gives a screen reader for the focused group.
PR build in Rancher (40 s): harvester-11-2 > Settings > upgrade-config > Edit; Tab to radiogroup "Restore VM", Tab to radiogroup "Shut Down VM on Migration Failure", Arrow Up to Enabled, Save (sent to that cluster with true, held); then harvester-118 (v1.9): the control is hidden and Restore VM is named.
main (f3ca9e4) on the same cluster: Restore VM's group has no name.
PR build on harvester-118 (v1.9.0-rc4): no new control.
What the new commits change, and what else I checked
Rebase:a88dde3 has the same content as the commit you reviewed; the only change is the feature-flags.js conflict, resolved by keeping main's componentHealth and adding this PR's flag after it.
New code: only a3392e7, two :aria-label lines and the test changes. Nothing new to raise in it, from either pass.
Still right: Save sends a boolean to the cluster you're on with the other fields kept; the control is hidden on v1.9 and the field isn't added; no page errors.
Docs:harvester/docs#1170 merged on 11-11. The v1.10 and v1.9 docs pages now explain the field, and the form's description links to the v1.10 one on a v1.10 cluster.
The two passes side by side
Cody's process
Ours
Verdict
Ready to approve (his flow leaves the verdict to you).
Approve, A-.
Earlier comments
2 of 2 Addressed.
2 of 2 Addressed.
How they were proven
AX tree of main and the PR in Rancher, a real Tab walk clip, axe; 7 breakages run against the new spec.
AX tree, a keyboard and Save clip, axe; 10 breakages run against the old and the new spec side by side.
Only this pass
Annotated before/after stills; checked the docs link now explains the field; the label 12036 gained.
Spotted that your review-body question is unanswered on GitHub; spotted that .105 now reports a master version; a UX point on unchanged code (S1).
Missed or wrong
Missed the review-body question. Its clip's title card said .105 runs v1.10.0-rc1; corrected and re-recorded before publishing. One breakage in its script hit the wrong handler first; fixed and re-run.
Nothing wrong found. S1 is real but on code you reviewed twice; see below.
Cost
In this session, with the earlier rounds in context.
About 18 minutes, 74 tool calls, about 223k tokens, as a fresh agent.
Pass B's evidence for the UX point it raised outside the new commits (see "Seen, not for the review").
Pass B's clip (26 s): keyboard to the named group, Enabled, Save on the v1.10 cluster, then the v1.9 cluster.
Seen, not for the review
UX, your call, I'd leave it out: the form still doesn't say what either value does. Disabled (the default) makes an upgrade wait with no timeout when a migration fails; Enabled leaves the VM off after the upgrade even with Restore VM on. The Storage Network, RWX Network and HTTP Proxy forms show a warning banner for their side effects, so the pattern exists. Why I'd leave it out: the code is unchanged since you reviewed it on 11-06 and 11-10; the author's first version had warning banners and removed them in round two "to reflect the latest design"; and the docs the form links to now explain both values with a caution. Pass B's comment text, if you want it: "Nothing on the form says what each value does to the upgrade. With Disabled, the default, the upgrade waits with no timeout while a VM's migration keeps failing. With Enabled, the VM stays off after the upgrade even when Restore VM is enabled. Other settings whose value affects the cluster show that on the form, like the warnings on HTTP Proxy and Storage Network."
Accessibility, shared component:aria-label also lands on each RadioGroup's outer div, because the shell's RadioGroup doesn't set inheritAttrs: false. axe marks the two wrappers aria-prohibited-attr as needs review, not a violation. The shell itself and this repo's IgnoreRuleModal.vue pass :aria-label to RadioGroup the same way, so it isn't the author's to fix; if anywhere, it belongs in rancher/dashboard.
Backend: issues 12036 and 12161 gained require/upgrade-v2 ("Require changes to Upgrade V2") on 11-12, a backend follow-up that doesn't touch this form.
.105: its API now reports master-9dd27fdb-head, which the gate treats as the newest version. The exact v1.10.0-rc1 string wasn't fed to the gate live this round; round three did that live, and Pass B checked it with a unit probe. The gate code hasn't changed since.
Not worth raising: swapping the new group's options to [false, true] with the labels unchanged passes every test, because the stub never pairs options with labels. Nobody writes that.
What I'd post
Approve, with one short body that also answers your own question. No thread replies needed: the body says both points are fixed, and you can resolve the two threads yourself.
Thanks, both points are fixed. On my question: I tried it as an extension in Rancher v2.16, on a v1.10 cluster where the control shows and saves, and on a v1.9 one, where it stays hidden.
No preview link, no code. If you'd rather keep the UX point, it would go in as a Comment instead of an approval, and it would be a fourth round for the author on code you've already passed twice.
What was and wasn't checked
Both of your 11-10 comments, live in Rancher on both clusters (accessibility tree, a real Tab walk, axe) and with breakages run against the old and new spec.
The rebase (range-diff against the commit you reviewed), the new commit's diff, CI on the new head (DCO, lint, unit-tests, Mergify all pass), and the PR conversation since 11-10.
The linked threads since 11-10: the docs PR merged, 12036 and 12161 relabelled, no new comments on 12148 or 12036.
Shared state: every save held in the browser, no admin session left open on mf, the shared admin's stored locale unchanged.
Baseline rebuilt before use: main f3ca9e4 = upstream tip, traefik, 200 with TLS verified. PR preview rebuilt at a3392e7.
Not checked: the standalone previews against a Harvester (VPN down); a real screen reader (the accessibility tree instead); a save reaching a backend; a real upgrade.