Akramin Akib

feat(apollo-wind): add ModelPicker for LLM Gateway model selection

feat(apollo-wind): WYSIWYG rich mode for PromptEditor Opt-in richText prop: formatting renders live while editing (RichTextPlugin + ListPlugin + markdown shortcuts) instead of raw markers with a separate preview. The PromptEditorToken[] contract is unchanged — rich state converts to/from markdown text tokens at the editor boundary: - utils/rich-serialization.ts: toolbar-scoped transformer set (bold/italic/ strike + lists; unknown markdown stays literal), shouldPreserveNewLines on both conversions so \n round-trips byte-identically, PUA-sentinel pill transformers so variable paths never pass through markdown parsing, a post-export unescape so literal * and _ (snake_case!) don't gain backslashes, and normalizeRichTextTokens (headless round-trip) for consumers' untouched-open commit guards - ToolbarActionsPlugin dispatches FORMAT_TEXT/INSERT_*_LIST commands in rich mode and reports active formats; toolbar buttons get aria-pressed states - Edit/Preview and mode='preview' are inert in rich mode (the editor IS the preview); MultilinePlugin is not mounted (it would flatten lists) - caret/drop/copy plugins generalized from paragraph-only guards to any non-inline block so pills inside list items keep working - 23-case round-trip guardrail corpus + 6 rich component tests

feat(apollo-wind): add PromptEditor extension points for host integration Additive API for hosts embedding the prompt editor: - validationOptions: validate chips against a wider set than autocomplete offers - index-normalized chip validation (records[0].id validates records[1].id), with normalizeVariablePath exported and free-form paths accepting indices - renderAutocompleteMenu: swap the built-in cmdk menu for a host picker - renderTokenPill: replace/decorate token chips (built-in pill passed through) - children: mount extra Lexical plugins inside the composer - strings: overridable UI strings for localizing hosts - showModeToggle=false + toolbarTrailing: widget-style toolbar (no Edit/Preview, formatting cluster left, host control right) - barrel re-exports for the helpers host plugins need ($insertTokenAtCursor, WORD_JOINER, createTokenNodeForOption, isPromptTokenNode, TokenPill, node classes, getEditorTokens/setEditorTokens)

fix(apollo-wind): make the ModelPicker trigger a combobox and keep the list behind the confirm Human review round on #1201. The trigger is `role="combobox"`, as wind's Select and Combobox triggers are. As a button labelled by `<label for>`, it announced as "Model, button": the label's asterisk is aria-hidden so required was silent, and the selected model is content, which a labelled button drops. A combobox's value *is* its content, and `aria-required` is valid on it, so both are announced now. The search field stays a second combobox for activedescendant, the same shape as wind's cmdk Combobox. Opening the delete confirm pulled focus into a portaled dialog, which Radix read as leaving the popover, so the list closed and focus landed on body. The popover now ignores close requests while a delete is pending, and the dialog hands focus back to the search field — the Material picker's behaviour. That exposed a second problem: a failed delete was routed through the same state as a catalog error, replacing the intact list with the message and painting the field invalid. Delete errors are an alert strip inside the popup now; catalog errors keep their treatment. Keyboard: Enter, Shift+Enter and Delete are gated the way the active descendant already was, so a refetch cannot select or act on a stale row. A collapsed section is one ↑/↓ stop (its header highlights) instead of being skipped, so → can reopen it in a mixed catalog, not only when every row is BYO. The grouped list scrolls the highlighted row into view; only the virtualized one did. A bare `value` with a BYO twin resolved to the BYO row because BYO sorts first. Hosts persist "no connection id" to mean the hosted model, so the hosted match wins when no connection id is given. Chip tooltips no longer render a <button>: the trigger is a span via `asChild`, so nothing inside `role="option"` is interactive, tabbable or form-submitting. `numericId` is dropped from `FolderSwitcherFolder` — nothing read it. Two docs still described a Lingui `i18n` instance. And the policy-blocked test reused a visible model name, so it passed with the filter removed; it uses a distinct name now.

fix(apollo-wind): keep ModelPicker keyboard focus inside the listbox model Review round on #1201. The substantive ones: The BYO row's edit/delete icons were tabbable buttons inside `role="option"`, which is an invalid child and pulls Tab away from the `aria-activedescendant` model the search field owns. They are `tabIndex={-1}` now, and the keyboard reaches the same actions from the search field: Shift+Enter edits the active BYO row, Delete on an empty search asks to delete it. Collapsing had two disagreements. The virtualized renderer honoured any key in `collapsedGroups` while the grouped one only collapsed BYO, so the arrow-key handler could hide a section in one renderer and not the other; both now share `isCollapsibleGroup`, as does the handler. And collapsing left `activeIndex` on a row no longer rendered, so `aria-activedescendant` named a missing element and Enter selected something invisible. The highlight now moves to the nearest visible row whenever its row hides, and Up/Down step over hidden rows. `defaultRowActions` recognised BYO by subscription type only, while grouping and tagging use `isByoModel`, so a row filed under Custom Models by its `byomDetails` got no actions. Same predicate everywhere now. A date-only `usageEndDate` was parsed as UTC midnight and formatted in the local zone, so anyone west of UTC saw the deprecation month one early. Date-only values format in UTC. Also: `ModelType`, `ModelFlatCosts`, `ModelTieredCost` were part of the `DiscoveryModel` shape but missing from both barrels; the `displayName` JSDoc referenced a `friendlyNameFor` prop that does not exist; the README badge walkthrough still described the deleted `i18n.ts`; and two stories promised behaviour they did not demonstrate.

fix(apollo-wind): localize remaining PromptEditor strings, review nits Review feedback: - token type labels (chip tooltip + autocomplete hint), the chip remove button's aria-label, and the empty-preview message join PromptEditorStrings; getPromptEditorTokenTypeLabel takes the labels - insertFreeForm becomes a '{path}' template rendered around the bold path, so verb-last languages can reorder it (a plain string is still treated as a prefix) - NODE_TYPE_TO_TOKEN_TYPE moves to shared/token-nodes, typed, replacing the PillRef cast in rich-serialization - ElementNode via a standard type import

feat(apollo-wind): add Underline and Code to the PromptEditor toolbar

fix(apollo-wind): preserve outer formats and pills around the new marks Review follow-ups on #1168: - UNDERLINE.replace built a fresh TextNode, dropping bits an outer transformer had already set, so `**<u>x</u>**` lost its bold on the next save. Carry node.getFormat() across. Underline still canonicalizes to the outermost position, which is stable and what the toolbar emits. - A pill enclosed by a code span was deleted. @lexical/markdown will not recurse into code-formatted nodes, so the sentinel never became a decorator and had no entry in the export-side table. Rescue those runs after import: the pill wins, the code formatting on that run is given up. - Toolbar wiring had no test; a wrong action name would have stayed green. Adds EditorToolbar.test.tsx driving every button against a stub actions ref. - Drops the separator left of the trailing slot.

refactor(apollo-wind): make ModelPicker presentation-only and field-like The picker now renders only what it is handed, and reads as a form field rather than a button. Strings: `i18n.ts` becomes `labels.ts`. Every user-visible string is a field on `ModelPickerLabels`, merged over English defaults, with the interpolated ones as functions so translators control word order. Two strings that were built in code — the context column and the section count — are fields now and localizable for the first time. Data: `requestContext`, `enableFolders` and `onModelDeleted` are gone and `models` is required. `usePlatformDiscoveryModels`, `useUserFolders`, `useCanManageByo`, `useDeleteByoConfiguration` and `useByoConnectionNames` stay exported as the composition kit. Two things that look like data stay in: the delete confirmation, because confirming a destructive action is presentation, and the `isBlockedByPolicy` filter, because a governance verdict must not be rendered no matter who fetched the row. `onEditModel` replaces the built-in AI Trust Layer navigation. Chrome: the trigger borrows `Input`'s classes verbatim so it sits flush with the text fields around it, with the invalid state driven by `aria-invalid` as `Input` does. Interactive controls opt into `cursor-pointer`, which Tailwind's preflight does not give buttons. Stories: six wrappers hardcoded Apollo's light-theme page grey, inherited from the apollo-react original. Under the dark themes that left the trigger on a light ground while its text followed the theme.

fix(apollo-wind): measure ModelPicker virtual rows and honour a changed groupBy Two items Copilot listed as previously missed on #1201. The virtualized renderer positioned every row at a fixed estimate with no `measureElement`, but rows are `min-h-*` and grow with a technical-id line, a BYO connection caption or host `renderMeta`, so neighbours could overlap or clip once the catalog crossed the virtualization threshold. Each rendered row now reports its real height to the virtualizer. `groupBy` was read once by `useState` although the hook's own option doc promised "pass a fresh groupBy to reset". A host changing the prop after mount now resets the view; the in-popup toggle still owns it between such changes.