migrate-radix-to-base

SKILL.md

Radix UI -> Base UI migration

You migrate shadcn wrappers, hand-rolled radix compositions, and their
consumers to @base-ui/react, keeping the project buildable at every step.
Be precise; never guess a mapping. When a prop or part is not in these
reference files, check node_modules/@base-ui/react/**/*.d.ts before
transforming, and record gaps in the report.

Preflight (always)

npx shadcn@latest info --json (or the project's runner): gives the
current base, STYLE (e.g. radix-lyra), tailwind version, aliases,
installed components, and package manager. Trust it over inference.

Detect the package manager (packageManager field / lockfile:
pnpm-lock.yaml, bun.lock, yarn.lock, package-lock.json) and use IT for
every install. Never leave a stale lockfile.

Require a clean git tree; work on a branch; one commit per component.

Baseline check BEFORE touching dependencies: run the project's
typecheck/build so pre-existing failures are never attributed to you.

Install @base-ui/react alongside radix. Radix packages are removed only
after the LAST component is migrated (both coexist fine).

Strategy: golden pair first, transformation engine second

Golden pair via the CLI (preferred). If the project is shadcn with a
known style (radix-<style>), the shadcn CLI itself is the golden-pair
executor:

Classify each ui wrapper FIRST: diff the user's file against its stock
origin, using the components.json style VERBATIM in the URL
(https://ui.shadcn.com/r/styles/<style>/<component>.json,
files[0].content). This works for prefixed styles (radix-nova) AND
legacy unprefixed ones (new-york, new-york-v4, default), which are all
still served.

WHOLE-PROJECT mode: flip components.json style radix-<style> ->
base-<style> now. PROGRESSIVE mode: do NOT flip yet (the project is
still mostly radix; the flip happens once, after the last component);
fetch base variants directly by URL instead
(https://ui.shadcn.com/r/styles/base-<style>/<component>.json).

PRISTINE wrappers, whole-project mode: shadcn add <component> --overwrite delivers the base variant with the project's exact
icon/font/preset resolution. Never bulk --all --overwrite; go
component by component, or you drown in unrelated registry version
drift. PROGRESSIVE mode: never use --overwrite (it destroys the
original that consumers still import); write the fetched base variant
content to <component>-base.tsx instead.

CUSTOMIZED wrappers: fetch the base variant and replay the user's diff
onto it (their customizations must SURVIVE; --overwrite would destroy
them). Mechanical implementation that works at scale:
git merge-file user.tsx radix-golden.tsx base-golden.tsx (three-way
merge, radix golden as ancestor) auto-resolves most files; hand-resolve
conflicts with the reference tables.

MANDATORY leftover sweep on EVERY golden-pair file, including ones that
merged "clean": grep -n "radix-ui\|@radix-ui\|IconPlaceholder" per
file. The registry sometimes reorders functions between variants, which
makes three-way merges report zero conflicts while leaving stale radix
hunks in place. A clean merge is NOT proof of a clean file.
This is more reliable than reconstructing transforms; use it whenever the
pair exists. Consumer/app code has no CLI mechanism: always hand-migrate it
against consumer-props.md.

Legacy styles (new-york, new-york-v4, default): classification only, no
replay. These have no base counterpart (there is no base-new-york), and
retargeting onto a base- variant would restyle the user's app. Use
the radix golden ONLY to detect customizations, then run the transformation
engine on the user's OWN file: rewire primitives, keep their exact classes,
apply class-mapping renames. Their look stays theirs. At the end of a
legacy whole-project migration, FLAG (do not fix): the style name still
reads as radix to the CLI, so future shadcn add will deliver radix
variants; the user decides whether to switch style or add manually.

Transformation engine (fallback). Hand-rolled radix code, non-shadcn
projects, unknown styles: transform using universal-patterns.md (imports
in BOTH forms: radix-ui and @radix-ui/react-*; asChild->render with the
worked example; Portal>Positioner>Popup; the positioner FORWARD rule; part
renames), the per-family props tables (overlays.md, menus.md,
form-controls.md, disclosure.md, display-misc.md), class-mapping.md
for data-attribute/CSS-var rewrites, and wrapper-shapes.md for exact
target shapes (tooltip arrow, SubContent defaults, select anatomy).

Modes

Progressive (default). "Migrate accordion" = one component, strangler-fig:

Detect in-progress state first: an existing <component>-base.tsx,
consumers split between old/new imports. The files ARE the state; resume,
never restart.

If the component imports other ui wrappers still on radix (select ->
button), STOP and recommend migrating those first, bottom-up.

Write the migrated version to <component>-base.tsx (original untouched;
golden-pair content fetched by URL, or transformed by hand, per the
strategy above); typecheck. Repoint consumers ONE AT A TIME (imports + the
call-site props in consumer-props.md); typecheck each. When no consumer
imports the original: delete it, rename -base -> original, flip imports
back, final check, commit. When the LAST radix wrapper in the project is
finalized, flip components.json to base-<style> and remove radix deps.

Whole project (only when explicitly asked): same per-component work in
dependency order (leaf/shared wrappers like button and label first). After
wrappers, sweep ALL app code against consumer-props.md — the call-site
break surface is much larger than asChild. Then remove radix deps, install,
full build.

Hard rules

NEVER touch non-radix libraries or their wrappers: cmdk (command), vaul
(drawer), sonner, input-otp, react-day-picker (calendar), recharts (chart).
Report them as intentionally untouched.

No Base UI counterpart: AspectRatio -> CSS aspect-ratio div; Label ->
native <label>; VisuallyHidden -> sr-only; Direction -> Direction
Provider (direction prop, not dir). Popover Anchor and NavigationMenu
Indicator have no equivalent: inert passthrough + flag.

button.tsx migrates to the REAL @base-ui/react/button primitive, never
a hand-rolled useRender wrapper.

Behavior deltas are FLAGGED, never silently patched (tabs manual
activation, menu items not closing on click, nav-menu 50ms delay). The
target is idiomatic Base UI matching the shadcn base registry.

Honest reporting: skipped/reverted files are listed as flagged, never as
migrated. Pre-existing failures are named as pre-existing.

Verify and report

Typecheck per file, build per batch, full build at the end vs the baseline.

Reports live in a .migration/ directory at the project root, ONE FILE PER
COMPONENT: .migration/<component>.md (e.g. .migration/accordion.md).
Rules:

Each run writes (or fully overwrites) the file for each component it
migrated. Re-running a component replaces its report; never touch other
components' files.

A multi-component run ("migrate alert-dialog and dropdown-menu") writes one
file per component, each self-contained; shared consumer-sweep notes are
repeated in every affected file.

Whole-project mode writes the per-component files plus
.migration/project.md (dependency swap, app-code sweep summary, final
build result).

There is NO index file. Migration status is derived from disk, not
maintained: scan the project's ui directory (the ui alias from shadcn
info, e.g. components/ui or src/components/ui) for remaining radix imports
when asked "what's left". End every run's summary with that derived count
("N wrappers remain on Radix").

Each .migration/<component>.md uses EXACTLY this structure (it is
documented publicly; reports must match it):

# <component>

<date, strategy used (golden pair via CLI / merge / engine), one-line verdict>

## Changed

<every file touched, with what changed and why; include file:line for
anything notable. Confirm the leftover scan is clean:
grep -n "radix-ui\|@radix-ui" on this component's files>

## Left alone

<files that look related but were intentionally not touched, with the reason
(cmdk/vaul/sonner are not radix; unrelated drift; etc.)>

## Behavior changes

<differences that compile fine but act differently; flagged, never patched
(tabs activation, menu close-on-click, delays...). Empty section if none>

## Verify by hand

<short manual QA checklist for this primitive family: focus return on
dialogs, keyboard nav + typeahead on menus/select, tooltip delay feel,
slider commit events. Concrete steps, one minute of clicking>

1f