✅ Done! Here's my read on all eight — code and tests look solid across the board. All the functional/unit CI is green; the only red on any of them is the PR-description gate, which is a description-template issue, not a code issue. ✅ Approve • #18215 — hook text. Clean. Picking the display field by

hook.type

(`command`/`prompt`/`system_prompt`) instead of first-non-empty is the right call, and the two new tests nail the exact mix-up case. No i18n surface. Ship it. • #18151 — button test IDs. Correct root cause:

BrandButton

reads

testId

, so the

data-testid

props were being swallowed. The rename in both modals is right and the tests verify the rendered attribute (they'd have failed on the old code). Nice catch. • #18144 — duplicate-name feedback. Good fix — the old code folded the duplicate check into

isNameValid

, so the field stayed

aria-invalid="false"

with no message. Splitting into `isNameFormatValid`/`isDuplicateName` and threading

error

into

SettingsInput

renders the

role="alert"

+ invalid styling correctly. Tests cover create and edit-on-own-name. One tiny nit (non-blocking): when

error

is set,

aria-describedby

switches from the format rule to the error id, so the rule text drops out of the accessible description — acceptable, just noting it. • #18112 — menu keyboard nav. Everything checks out: the initial-focus effect now keys on

!anchorElement || portalStyle !== undefined

so it fires once the portaled menu actually renders; arrow nav cycles

getEnabledItems

only (skipping the disabled "Set as active"); Escape/Tab restore focus to the

⋮

trigger before closing. Ports the #18019 fix cleanly and the 7 tests are genuinely behavioral. • #18111 — Default profile display. Right fix: normalizing

""

→

null

at the selector boundary keeps the internal

""

persistence semantics while rendering the Default item. Both entry paths (action-switch and selector

onChange

) covered. • #17968 — draft preservation. This is the most intricate one and it holds up. Capturing the input and conversation id in

pendingDraftRef

, guarding the debounce on the pending object identity, and moving the flush to a

useLayoutEffect

cleanup (so

innerText

is read before React detaches the node and loses block-element line breaks) are all correct. The tests cover rapid switching, home navigation, multiline, cleared-after-delivery, and conversation isolation. No notes. • #17967 — two-stage Escape. The

event.stopPropagation()

in the pill row's window

keydown

listener really does stop the modal's window listener (same target

window

, bubbling phase), so Escape scopes to the popover first — matches the two-stage test. The Enter/Space

stopPropagation

scoping to the trigger is a sensible bonus. • #17966 — header validation message. Reusing

validateEnvFormat

with a

errorKey

parameter is the right minimal refactor, and the new key is present in all 15 locales. Test asserts header-specific text shows and the env message doesn't. 🌱 Optional nits (not blockers) • #17967 / #17966 include

.pr/bugs/*.png

screenshots. The bot files an artifact-cleanup PR against

main

after merge for fork PRs, so it's handled — just be aware. • #17967 hardcodes expected pill text in the test rather than reusing the CSS-class constant pattern; harmless. ⚠️ One thing to fix before merge Every one of these fails

Validate PR description

on older runs, but they've since gone green (they fail only the required-review gate now). If you touch the descriptions again, keep the

## Why

/

## Summary

/

## How to Test

sections intact — that's the string the checker matches on. Want me to post these as actual GitHub reviews (approve #18215/#18151/#18144/#18112/#18111/#17968/#17967/#17966, or hold any back)? I didn't push anything yet.