chore(release): unblock the 0.1.0 publish pipeline

[rpvilo]

Gets the publish pipeline into a state where it can actually run. The headline is that the release workflow could not have succeeded as it stood — see the first item.

The blocker

.changeset/config.json still listed ignore: ["intentface"], the registry package deleted back in #44. Changesets treats an ignore entry that matches no workspace package as a fatal ValidationError, so changeset status, changeset version and changeset publish all aborted before doing anything:

ValidationError: The package or glob expression "intentface" is specified in the
`ignore` option but it is not found in the project.

Removed the entry. changeset status now resolves @intentface/chat at minor, and .changeset/README.md no longer refers to the deleted registry CLI either.

Changeset text

initial-release.md advertised a surface that no longer exists: StepQueue (removed in #30), ArtifactCard / ArtifactsPanel, and "styled components ship separately via the shadcn-compatible registry" (the registry was deleted in #44). Since this text becomes the published 0.1.0 changelog entry, it's rewritten to describe the eleven entry points that actually ship.

Workflows

Both were workflow_dispatch-only, commented "paused during initial development". Restored push: [main] + pull_request on ci.yml and push: [main] on release.yml.

Because this is the PR that turns CI on, it also has to leave main green — and main currently has two biome check failures, both pre-existing and unrelated to any of the above:

• packages/chat/tests/a11y/harness.a11y.test.tsx — the negative-control fixture trips useValidAriaValues on its deliberate aria-label="". Extended the ignore comment. (Same one-line fix as in feat(composer)!: remove TipTap engine and batch pre-1.0 renames #50; identical text, so the two merge cleanly.)
• components/chat.tsx — an unsorted import and an over-wrapped import block failing organizeImports. Applied Biome's own fix; reordering only, no semantic change.

Manifest

Added author and bugs to packages/chat/package.json. Deliberately did not add engines: the package targets browsers, and a Node floor there produces install warnings for consumers it doesn't actually constrain.

Verified

A real npm pack (not --dry-run, so the prepack/postpack export swap genuinely executes):

• Packed manifest's exports point at ./dist/src/*, and publishConfig.exports is consumed — the swap works
• All 11 entry points resolve to files that exist in the tarball
• 25 files, 74.5 kB: dist/, LICENSE, README.md and nothing else — no source, tests, or scripts leaked
• postpack restored the working tree to the ./src dev exports, leaving no package.json.prepack-bak

changeset version in a throwaway clone: bumps 0.0.0 -> 0.1.0 and generates a correctly attributed CHANGELOG.md. Worth knowing that @changesets/changelog-github hard-fails without a valid GITHUB_TOKEN (it escapes cleanly and changes nothing) — release.yml sets it at step level so both the action and the spawned bunx changeset version inherit it, which is correct.

Also green: 184 tests, tsc --noEmit, biome check, package build + publint.

Still needs you — not doable from here

1. Configure a trusted publisher on npmjs.com for @intentface/chat, pointing at this repo and release.yml. This can only be done after 0.1.0 exists on the registry — npm requires the package to already be published, both via the web UI and via npm trust. So 0.1.0 gets published manually; every release after that comes from CI with no credentials.
2. Flip the repo to public. Provenance is generated automatically under trusted publishing, but it requires a public source repository.

No NPM_TOKEN is needed — release.yml now authenticates via OIDC, so there is no long-lived secret to store or rotate. The @intentface org already exists.

Heads-up: npm/cli#8976 reports E404 on scoped packages published via changeset publish from changesets/action under OIDC — open since February 2026. That is this exact shape, so treat the first CI-driven release as unproven until it succeeds once.

Ordering

Merges cleanly with #50 (verified with git merge-tree). Either order works; both branch from main and target it independently.

---

Update: two release-blocking bugs in the built artifact

After the pipeline work above, I packed the tarball and installed it into a throwaway Next 16 app with no transpilePackages, so it resolved dist/ the way a real consumer does. That found two bugs that every green check we had — publint, tsc --noEmit, 184 tests, the app build — was structurally incapable of catching, because nothing in this repo has ever consumed the built output. The app compiles ./src via transpilePackages.

1. Every compound component shipped as unknown

bunup's default declaration emit uses TypeScript's isolated-declarations mode, which cannot infer the type of an Object.assign(Root, {...}) compound. It warned TS9010 and emitted:

declare const Thread: unknown;
declare const useThread: () => {};

All eight compound components — Thread, Composer, Message, Steps, Reasoning, Chip, Attachments, AskUser — plus several hooks and utilities were erased. The entire public API was unusable from TypeScript: JSX element type 'Thread' does not have any construct or call signatures.

Those TS9010 warnings were visible in the build all along, on main too, and I dismissed them earlier in this work as non-blocking because .d.ts files were still produced. That was wrong — they were produced with the types erased.

Fixed with dts: { inferTypes: true }, which routes declaration emit through tsc. reasoning.d.ts went from 568 B to 5.80 KB, which is the measure of how much was being dropped. publint passed before and after: it validates packaging, not type correctness.

2. The dev JSX runtime was bundled

dist/ imported jsxDEV from react/jsx-dev-runtime — 68 call sites, zero uses of the production react/jsx-runtime. Any consumer's production build throws:

TypeError: (0 , d.jsxDEV) is not a function

Bun's transpiler picks the runtime from the build process's NODE_ENV, and bunup was running with it unset. The jsx: { development: false } option looks like the declarative fix, but bunup 0.16.32 ignores it — verified, still 68 jsxDEV calls — so the build script pins NODE_ENV=production instead.

prepack also ran bare bunup, so it would have kept emitting the dev runtime even after build was fixed — and prepack is what produces the published tarball. It now calls bun run build, so there's one source of truth.

Verified against the packed tarball

With NODE_ENV explicitly unset in the shell, npm pack → install → next build:

• 15 react/jsx-runtime imports, zero jsxDEV
• Zero declare const … : unknown in any .d.ts
• A "use client" page using the compounds, their sub-components, useComposer / useReasoning / useThread, and event handlers — builds and prerenders
• A server component using /types, /message-utils, /chip-markdown — builds and prerenders
• A real consumer install pulls 29 packages: only @floating-ui/dom and nanoid beyond React

Also documented: compound sub-components and RSC

While isolating the above I hit a third thing, which is a genuine React constraint rather than a bug. Reaching a sub-component from a server component fails at runtime:

// Server component → "Element type is invalid… but got: undefined"
<Composer><Composer.Container /></Composer>

A server component importing a client module gets a proxy of that module's named exports; it cannot read properties off an exported value, and the sub-components live on the Composer object. Radix-style import * as Dialog sidesteps this because the namespace's properties are named exports — a compound object can't. Bare <AskUser /> works; <AskUser.Header /> does not.

This needs no code change — chat UI is interactive, so consumers will be in client components anyway — but the failure mode is cryptic enough to deserve documentation, so the README now has a React Server Components section stating the rule and showing both shapes.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
intentface-chat	Ready	Preview	Aug 3, 2026 3:30pm