mirror of
https://github.com/supabase/supabase.git
synced 2026-10-13 13:19:09 +03:00
chore/function-recent-errors
3
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
3e79df3ece |
feat(library): serve the block catalog as Markdown and harden the exporter (#50370)
## I have read the [CONTRIBUTING.md](https://github.com/supabase/supabase/blob/master/CONTRIBUTING.md) file. YES ## What kind of change does this PR introduce? Feature, bug fix. Part 3 of 6 in a stack that splits the library redesign into reviewable pieces. ## What is the current behavior? An agent can already fetch any guide as Markdown, but has no way to find out what guides exist: the entry point is a rendered React page. The exporter also fails quietly in ways that ship wrong output rather than failing the build: - An unknown component silently unwraps to its children, so a component rename drops its rendered content. - A registry item that cannot be read produces a page with no file listing. - A link to a missing page produces a 404 URL. - An unrecognized install framework produces a plausible command for the wrong CLI. - Only absolute `/library/docs` links are rewritten, so in-page anchors and sibling links break in the export. ## What is the new behavior? `/library` negotiates Markdown the same way the guides do — `Accept: text/markdown`, or an explicit `/library/index.md` — and returns a categorized catalog with every block, its description, its framework variants, and a link to each guide's Markdown. `config/library.ts` is the single catalog description the generator reads, and a test ties it to the content directory in both directions: a guide cannot be added without a catalog entry, or listed without a guide. Each quiet failure above now throws, and links resolve against the page they appear on and are checked against the set of published documents. ```bash curl -H 'Accept: text/markdown' https://supabase.com/library ``` ## Additional context `config/library.ts` also carries the category and preview metadata the redesigned homepage consumes in the last PR of the stack; here it is exercised by the Markdown index and its test. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit - **New Features** - Added a browsable library catalog covering categories, blocks, starter apps, and supported frameworks. - Added Markdown versions of the library homepage and documentation for compatible tools and workflows. - Added framework-aware links and expanded registry information, including dependencies and source details. - Markdown requests now work for the homepage and documentation, while browser requests continue receiving HTML. - **Bug Fixes** - Improved document link handling, metadata validation, slug consistency, and detection of duplicate or missing documentation entries. - **Tests** - Added coverage for catalog routes, Markdown generation, homepage negotiation, document parsing, and framework-specific links. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> Co-authored-by: Ivan Vasilov <vasilov.ivan@gmail.com> |
||
|
|
abbac3b852 |
refactor(library): resolve registry dependencies from one source of truth (#50367)
## I have read the [CONTRIBUTING.md](https://github.com/supabase/supabase/blob/master/CONTRIBUTING.md) file. YES ## What kind of change does this PR introduce? Refactor, bug fix. Part 1 of 6 in a stack that splits the library redesign into reviewable pieces. This one is the foundation the rest build on and has no visual change. ## What is the current behavior? Three build steps each reimplement "where does this registry file land in the user's project": `process-registry`'s `getDefaultPath`, `registry/utils`' `uniqBy` on `file.path`, and the Markdown exporter. They disagree, which produces real bugs: - A Vue block whose files come from `node_modules/@supabase/vue-blocks/` keeps its package path, so the installer writes the package folder into the user's project. - `registryItemAppend` builds its `docs` string from `(item.docs, items.flatMap(...))` — a comma expression, so the item's own docs are discarded. - A name collision between a block file and its client's file silently keeps one of the two. - Install commands guess the CLI family from substrings in the item name, so `infinite-query-composable` — a Vue block with neither "vue" nor "nuxtjs" in its name — gets the React CLI. - Production Vue installs use `@supabase/<name>`, but the `@supabase` namespace is registered with shadcn, not shadcn-vue. - `build:registry`, `build:content`, `build:markdown` and `build:llms` run in parallel, but the last three read `public/r`. ## What is the new behavior? `lib/registry-resolution.ts` owns installed-path derivation, first-party dependency naming, deduplication, and cycle detection, and every consumer calls it. `build-registry` validates the whole registry against shadcn's schema and resolves every item, so a broken reference fails the build instead of shipping. `clean-registry` throws rather than logging past a failure. Pages declare their install `framework` explicitly instead of it being inferred, and production Vue installs use the absolute registry URL. The build steps are serialized behind `build:prepare`, and a new `library-tests.yml` workflow runs the library's tests, checks the generated registry is committed, and builds the app. ## Additional context Regenerated registry artifacts are the mechanical result of the resolution fix — the Vue client items and the OAuth consent items that gained their client's docs. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **New Features** * Added explicit React and Vue framework selection for library blocks and installation commands. * Improved registry resolution, dependency handling, path validation, and Vue file normalization. * Added support for reliable local, preview, and production registry URLs. * **Documentation** * Updated Vue and Nuxt installation documentation to identify the Vue framework explicitly. * **Bug Fixes** * Preserved combined documentation and validated generated registry content more consistently. * **Tests** * Added coverage for installation commands, registry resolution, dependency handling, and generated artifacts. * **Chores** * Added automated pull-request checks for library tests and builds. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> Co-authored-by: Ivan Vasilov <vasilov.ivan@gmail.com> |
||
|
|
19d7233580 |
feat(ui-library): add headless app block for TanStack Start (#49579)
## I have read the [CONTRIBUTING.md](https://github.com/supabase/supabase/blob/master/CONTRIBUTING.md) file. YES ## What kind of change does this PR introduce? Feature — a new UI Library block. Stacked on #49573 (already in main) Fixes AI-1064 ## What is the new behavior? Adds `headless-app-tanstack`: customers sign in, authorize an MCP client, and use the product through agent tool calls. It composes the existing Password-Based Auth, OAuth Consent, and MCP Server blocks. - `/agents` provides a copyable connection prompt, lists OAuth authorizations, and lets customers revoke access. - The shared MCP runtime exposes `whoami` plus example task CRUD tools. Tools use the caller's Supabase client, with database grants and RLS enforcing ownership. - A root-level `supabase/` directory supplies local Auth/OAuth configuration, a declarative tasks schema, and Edge Function files, including `.env.example`. - Docs cover local setup, signing keys, migrations, environment configuration, deployment, and extending the tools. `/example/headless-app` previews the sign-in, consent, connect, and connected states. Shared block fixes make a fresh install work: - Explicit public URL resolution fixes OAuth discovery in local Edge Runtime when middleware runtime detection fails. Both external OAuth access tokens and ordinary authenticated app session tokens remain supported; embedded agents do not need an additional consent flow. - Registry targets keep backend files outside `src/`, and generated consumer routes omit source-only TypeScript suppressions. - Signup respects `auth.email.enable_confirmations`; sign-in/signup preserve the return destination. Missing consent IDs retain the existing error state without serializing `null` into the URL. ## How to test Use the UI Library on **staging** and follow the block pages' instructions. 1. Open the **Headless App** block page for TanStack Start. Install it into a fresh app and follow the setup instructions through connecting an MCP client. 2. Sign up, open `/agents`, and use the connection prompt to authorize a client. Call `whoami`, then create, list, update, and delete a task. 3. Confirm the client appears on `/agents`. Revoke access and verify it disappears and token refresh fails. An existing access token can continue working until it expires. 4. Follow the **MCP Server** block page's embedded-agent instructions using an authenticated app session. Confirm tools work without another OAuth consent flow and `whoami` returns `client_id: null`. 5. With a second user, confirm each user can only access their own tasks. Check that signup behaves correctly for the configured email-confirmation setting. 6. Check the Headless App preview states and run the installed app's typecheck and production build. ## Validation performed Fresh local installation and browser/SDK verification passed: 26 live MCP/Data API checks, 10 Deno tests, and 7 connection-page component tests. Also passed UI Library typecheck, targeted lint, registry/Markdown builds, and fresh consumer typecheck/production build. Both OAuth and ordinary app session authentication were exercised. Hosted deployment and consuming the confirmation-email link were not tested. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit - **New Features** - Added a TanStack Headless App example with sign-in, OAuth consent, MCP connection, and connected-agent screens. - Added task management tools for listing, creating, updating, and deleting tasks through MCP. - Added connected-agent management, including server URL and prompt copying, refresh, and access revocation. - Added a new Headless App registry block and documentation. - **Bug Fixes** - Preserved intended destinations through sign-up, email confirmation, and protected-route login redirects. - Improved OAuth discovery URL handling across forwarded-host deployments. - **Documentation** - Updated setup, environment, deployment, and Supabase CLI guidance for headless apps and MCP servers. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Saxon Fletcher <SaxonF@users.noreply.github.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com> Co-authored-by: repro <repro@local> Co-authored-by: Raúl Barroso <code@raulb.dev> |