Eight TDD tasks: the view-block scanner and virtual document, the HTML
service wrapper, merging HTML into completion and hover, folding and
linked editing, auto-close on type, standing down the duplicate client
provider, manifest guards, and a manual editor check.
Task 1 comes first because everything reads positions through it: its
length-and-newline invariant is what removes position mapping, and a
break there would misreport positions everywhere rather than fail.
The last task is manual verification in an Extension Development Host.
Unit tests cannot show that completions actually appear in an editor, and
a green suite has hidden non-functional features in this repo before.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Markup in a .wrn file highlights but has no tag or attribute completion,
no tag closing, and no tag-level folding: the grammar's embeddedLanguages
mapping only affects tokenization, and VS Code's HTML language service
never runs on these documents.
The design extracts view blocks into a virtual HTML document where
everything outside them is blanked to whitespace of identical length, so
source positions and virtual positions are the same and no mapping table
is needed. Region detection is a tolerant scanner rather than the parser,
because completion fires while the document is mid-edit and unparseable.
Completion merges WRNexus and HTML entries into one list ranked by
sortText, which also fixes an existing bug: the extension and the server
both answer completion on '<' today, so VS Code concatenates two lists.
Two decisions worth review:
- HTML formatting is excluded. formatWrn already formats markup, knows
WRNexus syntax, and would fight a second formatter that is free to
rewrite spacing inside @click={...} and client:visible.
- Only auto-close-on-type is client-side. Linked editing is standard LSP
and lives in the shared server.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
window.__wrnI18n was undefined in development: the payload shipped as an
executable inline script, and a document's CSP nonce is fixed at load, so
any such script arriving from a later response is blocked. Client
translations and language switching silently had no data.
The payload is now a type="application/json" block, which the browser
never executes and script-src therefore never applies to. The i18n
runtime, CSR navigation, and HMR all read the block instead of matching
window.__wrnI18n= with a regex.
Pages now render zero executable inline scripts, so an inline script-src
violation is structurally impossible rather than merely unobserved. Zero
framework JavaScript on island-free routes is unaffected: the block is
inert data, and nothing loads to read it unless the page needs it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The HMR client script was dead in the browser. HMR_CLIENT_JS is a
TypeScript template literal, so the regex [ \t\r\n] inside it was expanded
into real control characters, producing a regex literal containing a raw
newline — a syntax error that took the whole script down with "Invalid
regular expression: missing /". It now uses \s, and a test asserts the
emitted client parses and holds no control characters inside regex
literals; that test fails if the bug is reintroduced.
HMR also corrupted CSP nonces. A document's nonce is fixed at load, but
morph copied attributes from freshly fetched HTML, overwriting the live
nonce with one the browser will not honour. syncAttrs now leaves nonce
alone, and nodes moved across are re-stamped with the live nonce.
Islands vanished on every HMR update: morph puts the server placeholder
back over the mounted island. The island runtime now remounts on
wrnexus:hmr-updated. Remounting swaps the container for a bare clone —
re-rendering the existing root is a no-op once HMR has wiped the DOM
externally, and unmounting throws asynchronously because the nodes React
wants to remove are already gone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The island pieces existed but nothing connected .wrn compilation to island
emission. Now:
- codegen emits a data-wrn-island placeholder for component tags bound to
.tsx imports, keeping .wrn components on the normal mount path
- the dev pipeline and static build resolve island imports, thread the
names into codegen, and build the bundles
- collectScripts adds /__wrnexus/islands.js only when island markup is
present, so island-free pages still ship nothing
- island routes classify as static-interactive via hasIslands
Three bugs found by driving a real page in the browser:
1. The mount runtime was never built anywhere, so the bootstrap 404'd and
no island mounted.
2. Building the runtime separately from the islands gave each its own copy
of React: "Cannot read properties of null (reading 'useState')". The
runtime is now an entrypoint of the same build so React stays in one
shared chunk. The existing single-React test only compared bundles
within one build and could not see across build boundaries.
3. Island props arrived as attribute strings, so start={3} was "3" and
incrementing produced "31" then "311". Props now follow JSX semantics:
{…} parses as JSON, quoted values stay strings, and a runtime
expression is a WRN-ISLAND-PROPS build error rather than a silent
wrong value.
island-codegen.ts no longer imports @wrnexus/core. Compiler modules are
bundled into the Node-only VS Code extension, which contains no other
packages, so a runtime import of core broke the editor compiler; the two
helpers are implemented locally and the core dependency is dropped again.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two guards protect the core promise: a route with no islands emits no
assets at all, and a page with several islands keeps React in one shared
chunk.
buildIslands now writes a generated entry per island instead of passing
component sources directly. Two islands sharing a source deduped to a
single entrypoint, and output order is not guaranteed to match input
order, so island names could bind to the wrong bundle.
Island modules are excluded from the editor compiler bundle: it globs
packages/compiler/src, and island-bundle.ts calls Bun.build while
island-codegen.ts imports @wrnexus/core — neither belongs in a Node-only
VS Code artifact.
Integration assertions share one build. bun test interferes with
Bun.build's module reads after several build calls in one process, while
the same calls succeed repeatedly outside the runner; production is
unaffected.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Card, Carousel, Footer, Navbar, input, select, and textarea last changed
in d78707be, but the baseline was last regenerated several commits
earlier, so check:ui-visual already failed on main.
Unrelated to the React islands work; committed separately so the feature
changeset does not absorb it. Only source hashes changed — no UI source
was modified here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Implemented and removed. A render-phase flag cleared on a microtask is
still set when React runs effects, so islands writing from an effect —
the documented correct pattern — would throw. The flag also cannot be
set for an island's own re-renders, so real violations pass silently.
Detecting React's render phase reliably needs React internals, which is
not acceptable in a shipped framework. React already reports the real
hazard, and the getSnapshot caching requirement covers the loop case.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Twelve TDD tasks covering the approved spec: snapshot cache, store
bridge, marker codegen, .tsx resolution, error boundary, island runtime,
bundling with a shared React chunk, asset serving, route classification,
integration guards, the write-during-render guard, and HMR remount.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adds the approved design for consuming npm React components as
client-only islands inside WRNexus, without changing the SSR-first
rendering model.
Key decisions:
- New isolated @wrnexus/react package; react/react-dom as optional peers
- Client-only by default, SSR deferred to v2
- Islands declared via .tsx imports in .wrn frontmatter
- Store access via useSyncExternalStore, with the snapshot cache in the
adapter so @wrnexus/store stays untouched
- Bundling extends the existing Bun pipeline; React as a shared chunk
- Routes with no islands must still ship zero framework JavaScript
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The contract tracks a content hash per component source. Rewording the comment
in Typography.wrn changed its hash, so the check failed on an edit that alters
no rendered output. Regenerated intentionally; only that one entry moves.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The committed snapshot had drifted from source on main: dev-server exports
validateRpcCsrf, and styles exports the browser-cookie configuration types and
resolveBrowserCookieOptions. All six are additions, so the surface stays
backward compatible.
This unblocks check:production, which fails on main for the same reason.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Pagination reported the requested page only through its change output, so on a
server-rendered page the controls did nothing: the parent had to own client
state to react, and the emitted page never became a URL. An optional
hrefTemplate now renders the steps and page numbers as anchors, which work
before hydration and without JavaScript and give each page a crawlable URL.
Buttons remain the default for client-owned lists. End steps are clamped and
marked disabled rather than linking past the first or last page.
Output handler errors are no longer swallowed. Nothing awaits invokeOutput, so
a handler that threw became an unhandled rejection that never reached the
console and presented as a control that silently does nothing. Handler errors
are now reported as WRN-DEV-OUTPUT-HANDLER-ERROR.
Regenerates the component reference and the UI visual contract.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
It was written up in 4.7 but never made the work order, which is exactly how it
stayed dangerous in the first place.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three separate problems, all traceable to `git add -A` sweeping up a working
tree I had not inspected.
Commit 69020b25 ("docs: make the component sections executable") committed far
more than docs: 79 files of a half-scaffolded inter-app example, and four of
those files were truncated mid-statement. That broke `bun run typecheck` on
main. The example is reverted to its last green six-file form. The truncated
fragments and the fuller working copy are NOT in this commit -- if any of that
workspace was wanted, it needs to be reconstructed deliberately and committed on
its own, not as a side effect of a docs change.
Separately, `scripts/generate-ui-complete-catalog.mjs` was run while checking
which helper scripts still work. It rewrites components in place, so it
flattened six of them to stubs, deleted 24 more and lower-cased four filenames
before crashing. Contents were restored from HEAD, but the renames survived
that restore: Windows is case-insensitive, so `git status` reported clean while
Card, Container, Divider and Grid sat on disk under the wrong names. The index
now tracks the capitalised names, which is what the components declare and what
ui-redesign-contract.test.ts reads -- that test would have failed on any
case-sensitive checkout.
Documented both as 4.7 and 4.8 in the remediation plan, with the general rule:
no script that rewrites packages/ui/components/ may write in place. Also fixes
the heading level on 4.6, which was rendering outside section 4.
bun run check is green: 1,433 pass, 0 fail.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Expands 3.1 and 3.2 so the work can be done without re-deriving anything.
3.1 now records what 0.8.6 already fixed, separated into the ten components
that were miswired and the five that gained outputs they had been firing
undeclared, with the caveat that Map's three were converted but never confirmed
in a browser. For the 22 that remain it adds the finding that changes the
decision: all nine are pure scaffolds with no state, functions or handlers, and
five of them duplicate a component that already works -- FileUpload against
FileInput and FileUploadProgress, Toast and ToastNotifications against Toaster,
AdvancedDatePicker against DatePicker, AdvancedRangeSlider against RangeSlider.
Superseding those is a migration entry rather than new code, and leaves Chart,
TreeView, Confetti and CopyMarkup as the only ones needing to be built.
3.2 corrects the scaffold count from 23 to 28; the earlier figure used a looser
rule. Nine of the 28 are the 3.1 components, so the two items must be planned
together, and several of the rest are primitives that need only their styles
moved out of ui.css rather than any behaviour.
Also corrects the dead-output component count from 11 to 9 in both documents.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adds a per-subsystem measurement of reactive.js, made by minifying it
repeatedly with one subsystem removed rather than counting source bytes.
This corrects the earlier audit on both figures and on the conclusion drawn
from them. Component controllers are 23,722 bytes minified / 6,660 gzipped --
30.6% of transfer, not the "about 18%" previously claimed -- and splitting them
out saves 6.6 kB gzipped on a typical page, not "3-4 kB". Measured against the
example app, / and /login use none of the ten controllers and /layout uses one,
so most pages download and parse the lot for nothing.
The larger finding is that the runtime is not where the weight is. One page
parses 490,212 decoded bytes across 11 generated client modules while
transferring 21,026, and the largest module is 89.8% duplicated lines: the
state-restore prologue appears 162 times because client-codegen.ts inlines the
sync into every peer alias of every client function. Gzip hides it on the wire,
but parse cost follows decoded bytes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Collects what this session exposed into one actionable document: the five bugs
fixed in 0.8.6 and the guard protecting each, the places the model is
incomplete, the smaller defects, and the delivery and dev-loop problems.
Every item states the issue, the evidence, the change and a test that fails
before it. Where a cause is not proven -- the dev server not picking up
packages/ui edits -- the item says so and makes diagnosis step one rather than
asserting a fix.
Two standing conventions are written down at the top because the rest is
written against them: a component owns its markup, behaviour and styles in its
own .wrn file, and ui.css carries global styles only; and a test that still
passes with the fix removed is measuring nothing, which is how the 0.8.5 focus
trap shipped with no coverage.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An output only reaches a parent @binding when the component calls
output.<name>(). Two separate faults meant most of the library never got
there, and both failed silently at each end.
HTML lowercases attribute names, so a parent's @sizeChange registered under
"sizechange" while the component emitted "sizeChange". The lookup missed, fell
through to a DOM dispatch, and the binding was never invoked. That made all 17
camelCase outputs undeliverable -- DataTable.pageChange and .rowClick,
Map.markerClick, ChatBubble.messageClick, LayoutSplitter.sizeChange and the
rest. invokeComponentOutput now falls back to a case-insensitive lookup, and a
csr test fails without it.
Separately, 18 components dispatched hand-built CustomEvents rather than
calling output.*. A bubbling event on the component's own root never reaches a
binding, because parent handlers live in a registry only the output proxy
reads. Card, Footer, Breadcrumb, Accordion, alert, Badge, AnnouncementBar,
AvatarGroup, ToggleCount and InputNumber now emit properly; Marquee, Map,
Timeline, List and SearchBox additionally declare the outputs they were
already firing. Dispatches on window are left alone -- that is how Toaster,
Modal and DataTable signal across component boundaries.
Verified in a browser both ways before and after: an AnnouncementBar
dispatching its own bubbling "dismiss" never reached a page-level @dismiss,
and reached it immediately once it called output.dismiss().
This corrects the audit, which called the LayoutSplitter failure "narrow and
unexplained" and read 32 dead outputs as 16 components needing a rebuild.
"Outputs work elsewhere" was an assumption; the components that worked
happened to use lowercase names and output.*. The dead-output ratchet drops
from 32 to 22, and a new test forbids the raw-CustomEvent pattern outright.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A measured pass rather than a bulk rewrite. Numbers come from the source and
from a browser.
The migration entry covers what has accumulated since 0.8.5 and would
otherwise reach upgraders unannounced: the Tabs output contract, the Sidebar
BEM rename, the layout components leaving Tailwind so their rendered class
lists changed, LayoutSplitter and CustomScrollbar changing props and outputs,
the ui.css families that were removed, and the theme tokens that now paint
where they previously resolved to nothing.
The audit records what is still wrong, with counts: 32 outputs across 16
components that nothing emits, 23 components still on the scaffold pattern, 66
without a local style block and therefore dependent on ui.css, and 10 still
using Tailwind. A test pins the dead-output count at 32 as a ceiling that only
moves down, so rebuilding a component tightens it and no new one can be added
quietly.
It also records what is not worth doing. Splitting the runtime saves 3 to 4 kB
gzipped on a first visit to a file cached for a year, and hydration costs
1.5 ms for 21 scopes across 4325 elements, so neither is a real problem.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Section, SectionHeader and PublicPageShell move onto wire-* classes with
local style blocks, variants as data attributes. SectionHeader loses 38
utility lines, and Section stops spending one line per variant and colour
pairing: tinted and solid now select on two attributes. PageHeader was already
on the convention and only needed the audit.
examples/basic-app/app/pages/layout.wrn composes the whole set into one page.
Building it surfaced a library-wide bug. Ten custom properties were referenced
by components and defined by nothing: --wire-color-focus, --wire-color-surface-soft,
--wire-color-on-danger, --wire-color-surface-subtle and the input-* family,
plus hover and contrast for every semantic colour except primary and
secondary. An undefined custom property does not warn, it resolves to nothing,
so focus rings drew with no colour and every soft surface rendered
transparent -- 27 components referenced surface-soft alone. They are derived
in the theme now, and a test checks every token a component references against
the rendered theme CSS rather than the source, since most are generated.
The semantic spread also had to move ahead of the primary and secondary
entries so the palette keeps winning for those two.
Known and unresolved: LayoutSplitter emits its sizeChange output and the
component does fire it, but a parent binding on the tag is not invoked. The
tour page therefore points at the handle aria-valuenow rather than wiring a
handler that would never update. Outputs work elsewhere, so this is narrower
than an outputs-are-broken problem and needs its own investigation.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Container, Columns, Grid, Divider, Image, Link, Typography and Kbd were built
from utility classes and class: conditionals. That works only where Tailwind
is present, and every variant cost a dozen conditional lines -- Divider spent
eleven of them saying which token to paint the rule.
They now carry wire-* classes with a local style block, and variants are data
attributes the style block selects on. Divider went from eleven conditionals
to five rules, and Typography lost thirteen.
Behaviour is preserved rather than improved on. Container keeps columns and
gap even though a container is not really a grid, because applications depend
on them, and its columns default stays 2: the redesign contract test caught
that changing it would silently reflow every Container already published.
Additive only: Grid gains minItemWidth for an auto-fit track, Divider gains
dashed and dotted variants, Image gains fit, and Link gains underline.
Verified in a browser rather than by eye, since the pane cannot screenshot:
track counts match the declared columns at desktop and collapse correctly
below each breakpoint.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Both advertised behaviour they did not have. LayoutSplitter declared
resizeStart, resize and resizeEnd with no pointer handling whatsoever, so a
caller wired up @resize and received nothing, for ever, with no error, and its
props were columns, gap and maxWidth copied from a grid scaffold.
CustomScrollbar was the same shape with a scroll output.
The splitter now resizes. Dragging lives in the reactive runtime behind
data-wrn-splitter, because a pointermove fires far too often to route through
a client function and a state write made in that callback is dropped; the
resolved size is held on the container as a --wrn-split custom property and
the component grids from it. The handle is a real separator: arrow keys step
it, Home and End go to the bounds rather than to nothing, and it carries
aria-valuenow, aria-valuemin and aria-valuemax. minSize fixes both bounds so
neither pane can be dragged away and left unrecoverable.
CustomScrollbar is CSS rather than script -- scrollbar-width and
scrollbar-color with webkit rules for the engines that still need them -- and
its fake scroll output is removed rather than left unimplemented, since a
caller can listen for a plain scroll event.
The test harness needed a fix too: mount did not bind the window CustomEvent,
so the runtime built events from the host global and happy-dom listeners never
matched them, which made anything dispatched look silently lost.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The scroll lock was mine, and it broke every page carrying a Drawer or Modal.
Making dialog visibility testable, I replaced a size check with a data-show
check -- but a Drawer animates open, so its panel cannot be hidden with
data-show at all: display:none is not transitionable. Every closed Drawer
therefore looked open, took the body scroll lock and never released it, and
the page could not be scrolled. Both components publish data-open, which is
the signal that actually means open, and that is what is read now.
Stepper gains the wizard surface: showPanel renders each step body and shows
only the active one, the same contract Tabs uses, and controls adds Back,
Skip and Next, which becomes Finish on the last step. nextDisabled lets a form
hold the step; the component never validates anything itself, since the page
owns the form.
Stepper also gets a single root. The panels and controls were siblings of the
list, so the component had several roots and anything scoped to
data-ui-component missed most of it.
Tabs panels now slide in the direction of travel rather than fading upward.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Object props never worked in generated demos, and my two earlier attempts each
traded one failure for another:
- a bare {...} attribute is read by the compiler as an interpolation, so it
parsed JSON as JavaScript and the page 500ed
- parenthesising it compiled, but prop coercion runs JSON.parse on the raw
attribute, so ({...}) threw and every demo rendered empty and silent
- entity-escaping the braces did not help either: the compiler hands the
attribute over without decoding, so JSON.parse still failed
They are now hoisted into page state and bound, which is what the playground
has always done. The state initialiser uses JSON.parse rather than an object
literal because the parser reads a leading brace as the start of a block.
Navbar gains a real profile: a brand, links, a two-column dropdown panel and
calls to action, instead of the generic scaffold samples that made every demo
look identical and showed no dropdown at all.
MegaMenu closed while the pointer travelled to it. The panel sits below the
trigger and that offset belongs to neither element, so crossing it fired
mouseleave on the root. A descendant now covers the gap.
Scrollspy could not be exercised at all: its links pointed at ids that did not
exist on the page. The demo now ships real sections, in page flow because the
runtime observes against the viewport.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two orphaned controllers in the reactive runtime targeted markup nothing
emits any more: hydrateSidebarControllers looked for .wire-sidebar-shell and
friends, which the Sidebar rewrite replaced with BEM classes earlier today,
and hydrateDropdownControllers looked for [data-wrn-dropdown], which no
component or compiler output has ever produced.
ui.css loses the matching legacy sidebar rules, the wire-mega-menu family
left behind when the MegaMenu scaffold was replaced, and a set of
self-contained application-pattern families that nothing references.
Utility layers are deliberately kept even where an individual member is not
name-checked anywhere. wire-bg-primary is documented and tested while
wire-bg-secondary is not, but they are one public family and splitting them
would be incoherent.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The runtime budgets measured raw source, which counts comments -- and the
production build minifies, so comments cost a visitor nothing. The metric
therefore rewarded deleting explanatory comments over writing smaller code,
and could not tell a real feature from a wall of prose.
They now measure the minified output, which is what is actually served:
/__wrnexus/reactive.js is its own file, minified, with an immutable year-long
cache. The reactive runtime is 69684 minified against a 80000 budget, from
175246 raw -- roughly 21kB gzipped, fetched once.
Also fixes two real bugs found while testing the showcase:
- object-valued props were serialised as a bare {...} attribute, which the
compiler read as interpolation and tried to parse as JavaScript. That
returned 500 for /components/navbar. Arrays start with [ and were never
affected, which is why only object props broke. All 108 pages now render.
- the runtime walked text nodes inside textarea, script and style, so a
JSON sample in a textarea was evaluated away.
Navbar styles move out of ui.css into the component, matching the rest of the
navigation group. No declarations changed: 4519 before and after.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Scrollspy replaces a scaffold that rendered bare anchors. The runtime observes
the sections the links point at and writes the marker straight onto the links:
an IntersectionObserver callback fires long after the client function that
registered it returned, so a state write there would be dropped.
Navbar and Breadcrumb both emitted aria-current="" for every inactive link.
That is not a valid value -- the attribute takes a token or must be absent --
so every link claimed a state it did not have. Breadcrumb had it too, despite
being the strongest component in the group.
Navbar also takes roving arrow-key focus across its menu bar.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Clintchiz
