feat(forms-rjsf): accept @rjsf 6.x and make the theme injectable

[thorwhalen]

Closes #57.

The problem

@rjsf/shadcn is published on the 6.x line only (6.8.0, peering @rjsf/core@^6.8.0). acture-forms-rjsf@1.0.0 peered @rjsf/core@^5.20.0. The two could not be installed together — an npm peer conflict at install time, i.e. before anyone wrote a line of form code. The adapter was unusable by exactly the consumer it was written for.

There was a second half to that, not named in the issue: even with the range fixed, there was no way to hand the theme to the component. rjsf-form.tsx imported Form from @rjsf/core and rendered it directly. The README's "Theming" section told readers to "pass your own Form from the themed package" and no prop accepted one.

What changed

1. Peer range → ^5.20.0 || ^6.0.0 for @rjsf/core, @rjsf/utils, @rjsf/validator-ajv8. Widened, not moved to 6-only — see Why dual below.

2. <RjsfForm /> takes an optional form prop — ComponentType<FormProps>, which is the type every RJSF theme's default export already has — defaulting to @rjsf/core's unstyled Form. Also exports the RjsfFormComponent type.

import ShadcnForm from '@rjsf/shadcn';
<RjsfForm form={ShadcnForm} command={cmd} onSubmit={run} onCancel={close} />;

This is hard-don't #8's prescribed fix ("Adapter packages expose unstyled components OR a config slot"), not a violation of it: @rjsf/shadcn is a devDependency for the smoke test only — never a runtime dependency, never a peer. No UI kit is bundled. That claim now has a test (below); it did not before.

3. Both halves of the declared peer range run in CI — a new rjsf5 job. See The 5.x matrix below.

4. Docs updated in the same pass — package README (install + a rewritten Theming section with the palette-adapter binding and the form-stability note), the index.ts module docstring, the npm description, and the acture-palette-design skill's Form adapters section.

Adapter code needed no change to span the two majors

Checked at source, from the published .d.ts of @rjsf/core@5.24.13 and @rjsf/core@6.8.0. Everything this adapter touches is shape-identical:

	5.24.13	6.8.0
@rjsf/core default export	Form	Form
FormProps exported	yes	yes
@rjsf/validator-ajv8 default export	validator instance	validator instance
schema / validator / formData / children	same	same
onSubmit	(data: IChangeEvent, event) => void	identical
liveValidate	boolean	'onChange' | 'onBlur' | boolean

The only signature that moved widened, so the false we pass is still valid. @rjsf/utils is a declared peer but is never imported by this package's source.

Why dual rather than 6-only

• The repo's own precedent: feat(ai-vercel)!: require AI SDK v5+ and project tools onto inputSchema #55 moved acture-ai-vercel's ai peer to ^5 || ^6 || ^7 while dev-testing only the newest major.
• Widening is a minor; narrowing to ^6.0.0 would be a breaking major that the issue did not ask for as its primary option, and it would strand any 5.x consumer for no gain.
• The code genuinely honours both (table above) — and both are now checked on every PR, which is the difference between this and feat(ai-vercel)!: require AI SDK v5+ and project tools onto inputSchema #55's precedent.

The 5.x matrix (review repair — the first draft of this PR got this wrong)

The first draft shipped a hand-run-once 5.x check plus an argument that CI could not do better. The argument was overstated and the gap it left was real. Both are now fixed.

The gap, demonstrated. On the previous commit, change rjsf-form.tsx's liveValidate={false} to liveValidate="onChange" — a legal, documented 6.x value (the widening in the table above). pnpm typecheck exits 0, all tests pass, CI is fully green, and the package publishes with "@rjsf/core": "^5.20.0 || ^6.0.0". The same source against real @rjsf/core@5.24.13 declarations:

src/rjsf-form.tsx(78,9): error TS2769: No overload matches this call.
    Type 'string' is not assignable to type 'boolean | undefined'.

A consumer who installs acture-forms-rjsf alongside @rjsf/core@5 — which the peer range invites — gets a hard compile error, and nothing in this repo could have caught it. Acceptance criterion 2 of #57 was unmet.

Where the impossibility argument went wrong. The premise is true: pnpm matches peer dependencies by package name, so an npm:-aliased 5.x @rjsf/core silently binds the 6.x @rjsf/utils, and packageExtensions is not honoured by pnpm 11.1.1 from either pnpm-workspace.yaml or package.json#pnpm. But that rules out two alias trees inside one install. A CI matrix is two installs, and the second one needs no aliases at all.

What was built. scripts/pin-rjsf-5x.mjs rewrites the three @rjsf devDependencies to the 5.x clause read off the declared peer range, drops the 6.x-only @rjsf/shadcn and its smoke test, and then a plain pnpm install --no-frozen-lockfile resolves a plain 5.x tree. The rjsf5 CI job runs it, then --verify (a job that silently tests the wrong major is this same bug again), then builds acture, typechecks and tests. Verified end to end in a throwaway worktree:

pin-rjsf-5x: pinned @rjsf/core, @rjsf/utils, @rjsf/validator-ajv8 to "^5.20.0", dropped @rjsf/shadcn and its smoke test.
pin-rjsf-5x --verify: @rjsf/core@5.24.13 — a real 5.x tree.
$ tsc --noEmit
 ✓ src/rjsf-form.test.tsx (4 tests) 72ms
      Tests  4 passed (4)

…and red on the mutation that motivated it:

$ (liveValidate="onChange")  →  src/rjsf-form.tsx(87,9): error TS2769 … Exit status 2

The script reads the 5.x clause rather than hardcoding it, so narrowing the published range to ^6.0.0 makes the job no-op instead of fail — the matrix retires together with the promise it exists to check (verified: both pin and --verify exit 0 with "no 5.x clause. Nothing to check — skipping.").

The one thing that genuinely stays 6.x-only is the @rjsf/shadcn smoke test, because the theme has no 5.x release. That is why it lives in its own file.

Guards and their mutation tests

Every number below re-measured on the current branch state (17 tests in packages/forms-rjsf).

Mutation	Result
Narrow peerDependencies to ^6.0.0	3 failed | 14 passed
Declared range drifts off the installed major (^5.20.0 || ^7.0.0)	3 failed | 14 passed
Adapter ignores the injected form (form ?? Form → Form)	2 failed | 15 passed
@rjsf/shadcn promoted to dependencies	2 failed | 15 passed
A workspace package declares a tailwindcss peer	1 failed | 16 passed
ignoreMissing: [tailwindcss] deleted from pnpm-workspace.yaml	1 failed | 15 passed | 1 skipped
(restored)	17 passed

Three of those guards are new in the repair commit, and each closes a claim this PR made with nothing behind it:

• Hard-don't feat: Phase 4 (v1.0) + v1.1 — tier system, CLI, devtools, snapshot, tier warnings #8 had no test. Moving "@rjsf/shadcn" from devDependencies to dependencies — the obvious "make the theme just work for consumers" edit — pulled radix + lucide-react + tailwind-merge + tailwindcss-animate into every consumer's install while all 11 tests stayed green. Now asserted absent from dependencies and peerDependencies, plus the stronger form the sibling adapters already satisfy: this package declares no runtime dependencies at all.
• The tailwindcss suppression had no compensating check. peerDependencyRules.ignoreMissing: [tailwindcss] is workspace-wide and permanent, and pnpm honours no scoped form (tailwindcss-animate>tailwindcss and @rjsf/shadcn>tailwindcss were both tried; both still report the unmet peer). So the day a workspace package genuinely peers on tailwindcss, its missing peer would go unreported. The check asserts no workspace package declares that peer while the rule is in place — and asserts the rule is still there, so the block retires with it.
• One of the two theme tests did not guard the theme. "submits through the theme" passed on an adapter that ignored form entirely, because @rjsf/core's own form submits identically. Both tests now assert theme identity through one expectThemedInput helper, keyed on @rjsf/core's form-control class (a discriminator that survives Tailwind class churn) rather than on rounded-md alone. The form ?? Form → Form mutation went from 1 failed to 2 failed.

The form prop must be referentially stable

FormImpl is recomputed on every render, so an inline theme — form={withTheme(myTheme)} in JSX, or an un-hoisted adapter arrow — is a new component type each render. React reconciles by element type, so the form unmounts, remounts, and whatever the user had typed is silently discarded and reset to defaults.

Memoizing cannot fix this, which is worth stating because it is the obvious remedy: the changed identity arrives as the prop, so useMemo(() => form ?? Form, [form]) invalidates on exactly the renders that matter. The requirement is documented instead — on the prop's JSDoc, in the README's Theming section (both snippets there are safe precisely because they bind at module level), and in the acture-palette-design skill.

Test delta

	before this PR	first draft	now
acture-forms-rjsf	3	11	17
workspace	587 / 21 packages	595 / 21 packages	601 / 21 packages

pnpm build → 0, pnpm typecheck → 0, pnpm test → 0 (exit code checked, not inferred). pnpm install --frozen-lockfile clean. pnpm peers check → No peer dependency issues found.

packages/forms-rjsf test:  Test Files  2 passed (2)
packages/forms-rjsf test:       Tests  17 passed (17)
TOTAL: 601 tests passed across 21 packages, 0 failed

One incidental

pnpm peers check was clean before this PR and is clean after. Keeping it that way needed peerDependencyRules.ignoreMissing: [tailwindcss] in pnpm-workspace.yaml: @rjsf/shadcn pulls tailwindcss-animate@1.0.7, whose peer range is the malformed ">=3.0.0 || insiders". insiders is not a semver range, so pnpm treats the whole range as unsatisfiable and reports the peer unmet no matter which tailwindcss is installed — verified by installing tailwindcss@4.3.3, which did not silence it. Nothing here compiles CSS. Commented in place, and now covered by the compensating test above.

Deliberately not done

• The rjsf5 job is on ci.yml only, not on release.yml's gate. ci.yml also runs on pushes to main, so the 5.x check does run alongside every release; it is just not a needs: edge on the publish step. Worth revisiting if a release ever lands without a green ci run.
• No roadmap entry. Tracked by the issue + changeset + CHANGELOG; the comparable fix PRs (fix(mcp,hotkeys): harden v1.14–v1.18 surface before reelee-web wiring #51, fix(core): drop NUL-byte separator from the when-clause regex tokenizer #53, feat(ai-vercel)!: require AI SDK v5+ and project tools onto inputSchema #55) did not touch docs/roadmap.md either.
• No uiSchema / validator props. Neither is needed by Widen acture-forms-rjsf's @rjsf/* peer range to 6.x — @rjsf/shadcn ships on 6 only, so shadcn consumers cannot install it #57, and PaletteFormAdapterProps carries no uiSchema for the palette path to supply.
• No createRjsfFormAdapter factory. A one-line arrow at the call site does it.

minor changeset: .changeset/rjsf-6-peer-range-and-theme-prop.md.

https://claude.ai/code/session_0191cTtmnhFbC2ZKNtgxHXye