mirror of
https://github.com/supabase/supabase.git
synced 2026-09-06 09:59:03 +08:00
create-pull-request/patch
6 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Miranda Limonczenko
|
f10f00ae69 |
fix(e2e): install e2e-shared when CI filters to a single suite (#48960)
## 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? Bug fix. Unblocks the WWW E2E check on `master`. ## What is the current behavior? The WWW E2E job fails before running any test: ``` Error: Cannot find package '@axe-core/playwright' imported from /home/runner/_work/supabase/supabase/e2e/shared/axe.ts Error: No tests found ``` Both E2E workflows install with a filter: ``` pnpm install --frozen-lockfile --filter=e2e-www... ``` The `...` suffix pulls in a package's declared dependencies. Neither `e2e-www` nor `e2e-docs` declared `e2e-shared`; both reach it through relative imports such as `../../shared/axe.ts`, which pnpm's dependency graph cannot see. So the filter selected one project, `e2e/shared/node_modules` was never created, and Node resolving `@axe-core/playwright` from `e2e/shared/axe.ts` walked up to a root that does not carry it under pnpm's isolated layout. `e2e-docs` is broken the same way. It had not run against the shared module yet, so it has not gone red. ## What is the new behavior? `e2e-shared` is declared as a workspace dependency of both suites, so the filter installs it. | | Filter scope | Importing `e2e/shared/axe.ts` | | --- | --- | --- | | Before | 1 of 28 projects | `Cannot find package '@axe-core/playwright'` | | After | 2 of 28 projects | Imports cleanly | The lockfile gains two `link:../shared` entries and no new downloads. ## Manual Testing 1. Check out this branch and delete the shared package's modules: `rm -rf e2e/shared/node_modules` 2. Run the command CI runs: `pnpm install --frozen-lockfile --filter=e2e-www...` 3. Confirm the output reports `Scope: 2 of 28 workspace projects` and that `e2e/shared/node_modules` exists again. 4. Repeat steps 1 - 3 with `--filter=e2e-docs...`. ## Additional context Fixing only the workflow lines, by adding a second `--filter=e2e-shared`, would work as well. Declaring the dependency was chosen instead because the dependency is real and every consumer of the filter gets it, not just the two workflow files. The imports stay relative. Declaring the workspace dependency is enough to get the package installed, so no import paths change in this PR. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **Chores** * Updated end-to-end test packages to use shared testing utilities at runtime. * Improved consistency between documentation and website test suites. <!-- end of auto-generated comment: release notes by coderabbit.ai --> Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
Miranda Limonczenko
|
777c02c205 |
test(docs): scan changed pages for WCAG 2.1 A/AA in warn mode (#48727)
Closes DOCS-1233 ## 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? Test coverage. The docs accessibility check now covers the full WCAG 2.1 A/AA rule set instead of two rules. **Note:** This PR tests _only_ the main article of changed pages (meaning, the content itself). A follow-up Linear issue is to address scanning the pieces outside of that: header, navigation, and interactive elements. ## What is the current behavior? The `@a11y` test in `e2e/docs` runs two axe rules against each in-scope page, `heading-order` and `page-has-heading-one`. Both already pass everywhere, so the check only guards a result we have. Nothing else in WCAG A/AA is checked. ## What is the new behavior? The same test runs the full WCAG 2.1 A/AA rule set. - **Existing debt does not block PRs.** Only the two heading rules fail. Everything else reports. - **The check stays fast.** It scans the article only and skips nine rules that cannot fire there. Scan time drops from 2405ms to 981ms. - **Findings belong to us.** Legacy mode excludes cross-origin frames. YouTube embeds were counting against us, 11 of 15 violations on one page. - **A pass carries meaning.** A 404 reports as a load failure, not an a11y bug. A page scanned before it hydrates warns instead of quietly reporting clean. ## How the findings appear The test is named `has no blocking accessibility violations`, so a failure listed by CI is always something to fix. It is not named for the full rule set, because a green check would then claim more than the check verifies. | | Rules | Where you see it | | --- | --- | --- | | Blocking | `heading-order`, `page-has-heading-one` | Test failure, so the runner reports it on the PR | | Reported | Everything else in WCAG A/AA | `::warning` annotation on the run | An annotation looks like this, on a run that still passes: ``` ::warning title=Accessibility::/docs/guides/database/functions has 1 non-blocking accessibility finding(s): frame-title (4) ``` The full axe result for each page is attached to the report as `axe-results.json`. ## Matching the Studio ratchet This follows the ESLint ratchet in `apps/studio`. That pattern warns on pre-existing debt rather than blocking on it, surfaces findings as annotations rather than PR comments, and promotes a rule to an error once its violations reach zero. The mechanism here is `ENFORCED_RULES` in `utils/axe-helpers.ts`. The two heading rules are on it because the heading-hierarchy work drove them to zero site-wide. The intent is to migrate rules into that list one at a time. Pick a rule, fix its violations, then move it into `ENFORCED_RULES` so it cannot come back. An exhaustive scan of the site groups the current backlog by root cause to sequence that work, and two fixes cover 99.1% of it. Studio keeps per-file baseline counts, which this does not. A whole-rule list is coarser, and it works here because docs violations reach zero across the site rather than per file. ## Manual testing Install the browser once, then run each step from the repo root. Every command scans production, so you do not need a local docs server. ```bash pnpm -C e2e/docs exec playwright install chromium ``` 1. Confirm a reported finding does not fail the check. ```bash DOCS_E2E_PAGE_PATHS=/docs/guides/database/functions PLAYWRIGHT_BASE_URL=https://supabase.com pnpm e2e:docs:a11y ``` Expect `1 passed`, and the `::warning` annotation above in the output. 2. Confirm the scan finds that violation. Same page, now failing on every rule. ```bash A11Y_ENFORCE_ALL=1 DOCS_E2E_PAGE_PATHS=/docs/guides/database/functions PLAYWRIGHT_BASE_URL=https://supabase.com pnpm e2e:docs:a11y ``` Expect `1 failed`, reporting `frame-title (serious, 4 node(s))`. Steps 1 and 2 together are the point of this PR. 3. Confirm the skipped rules stay skipped. ```bash A11Y_ENFORCE_ALL=1 DOCS_E2E_PAGE_PATHS=/docs/guides/getting-started/quickstarts/nextjs PLAYWRIGHT_BASE_URL=https://supabase.com pnpm e2e:docs:a11y ``` Expect `button-name (critical, 2 node(s))` and `label (critical, 2 node(s))`, and no `color-contrast`. 4. Confirm a page that does not load reports a load failure. ```bash DOCS_E2E_PAGE_PATHS=/docs/guides/does-not-exist-xyz PLAYWRIGHT_BASE_URL=https://supabase.com pnpm e2e:docs:a11y ``` Expect `Expected a successful response for /docs/guides/does-not-exist-xyz, got 404`, and no axe assertion. 5. Confirm the link checker still passes alongside the a11y test. ```bash DOCS_E2E_PAGE_PATHS=/docs/guides/auth/passwords PLAYWRIGHT_BASE_URL=https://supabase.com pnpm e2e:docs ``` Expect `3 passed`. ## Known gaps - `/docs/reference/*` is not scanned. Those routes render client-side into tens of thousands of elements, where axe exceeds its timeout and results depend on whether the scan caught the page mid-render. - Shared chrome is outside the article scope, so nav, sidebar, footer, menus, and drawers are not covered. - axe catches roughly 30-40% of WCAG issues. Keyboard navigation, focus management, and screen reader behavior still need manual testing. --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
Miranda Limonczenko
|
8dda0c3910 |
Add heading-hierarchy a11y check to docs E2E tests (#48422)
Closes DOCS-1232 ## Problem We do not have any tests to verify that we are following a proper heading hierarchy. For a documentation site that deals in mostly static content, this test is important. Single h1 + logical heading hierarchy (h1→h2→h3, no skips) matters because screen reader users navigate by jumping between headings — broken structure breaks that navigation. Relevant: WCAG 1.3.1 Info and Relationships (Level A) — https://www.w3.org/WAI/WCAG21/Understanding/info-and-relationships.html ## Solution Add Playwright axe-core, which we plan to expand later, to test only the h1 and header-hierarchy rule. This is added to our current suite that dynamically checks only pages that are edited. ## Manual testing 1. Find a docs guide and intentionally break the header hierarchy. 2. Run `pnpm e2e:docs:a11y` and see your errors. 3. Resolve the issue and run again to see errors resolved. Ensure there is at least a line changed to see the page tested. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **Tests** * Added automated accessibility checks for documentation pages. * Verified heading order and the presence of a level-one heading on each page. * Added a dedicated command to run documentation accessibility tests. <!-- end of auto-generated comment: release notes by coderabbit.ai --> Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> |
||
|
Miranda Limonczenko
|
52cb1c2600 |
feat(docs) Dynamically E2E test all docs-owned content (#48320)
Closes DOCS-1203 ## Problem The docs E2E workflow only ever tested one hardcoded page: the Next.js quickstart. All other docs content had no E2E coverage. ## Solution This PR expands the initial scaffolding to generalize the Next.js quickstart tests, page runs and checks local links, to all pages affecting Docs content: - Add `resolveDocsScope` (`e2e/docs/utils/resolve-docs-scope.ts`) to map changed guide and troubleshooting `.mdx` files to their `/docs/...` page paths, and to expand changed `_partials` to every page that includes them (including transitively, through partials nested inside other partials). Federated guide sections (`graphql`, `database/extensions/wrappers`, `ai/python`, `deployment/terraform`, `deployment/ci`) and reference docs stay out of scope, and resolution is capped at 20 pages to keep runtime bounded. - Replace the single `quickstarts.spec.ts` test with a generic `docs-pages.spec.ts` that loads whatever pages are resolved, asserting each renders with an `<h1>` and that its docs-owned links resolve. - Add `run-e2e-docs.ts` so `pnpm e2e:docs` resolves scope locally (from commits since `origin/master`, plus staged/unstaged changes) and skips Playwright entirely when nothing in scope changed. - Update `.github/workflows/docs-e2e.yml` to widen the trigger paths to all guides/troubleshooting/partials, resolve scope in a dedicated step, skip the rest of the job when scope is empty, and accept a `page_paths` input for manual `workflow_dispatch` runs. - Rewrite `e2e/docs/README.md` to document the new scoping behavior, the override envs (`DOCS_E2E_PAGE_PATHS`, `DOCS_E2E_BASE_REF`), and how CI uses the suite. - `pnpm e2e:docs:all` is also added to run tests on every page locally. Good for scoping issues but should not be included in CI. ## Manual testing Walk through the following steps to verify this works: - [x] `pnpm e2e:docs` from repo root resolves the expected pages for a local guide edit and can run against local dev **Note:** Challenges with testing on local in part because of the long lag for first page load. Recommendation to use a hosted URL is added to docs. - [x] Editing a shared `_partials` file resolves to every page that includes it (including through nested partials) - [x] `pnpm e2e:docs` exits cleanly with no Playwright run when no in-scope files changed - [x] `git diff --name-only ... | pnpm -C e2e/docs resolve-docs-scope` prints the expected page list for a sample diff - [x] Workflow run on a PR that only touches `e2e/docs`/workflow files skips the Playwright steps - [x] Manual `workflow_dispatch` run with `page_paths` set tests only those pages - [x] Run `pnpm e2e:docs:all` to run the suite on all docs content, which takes awhile ## Next steps After this PR merges, we have the scaffolding to add more fun tests like a11y 😁 <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit * **New Features** * Added scoped Docs E2E runs that target eligible doc pages based on changes, plus manual page-targeted runs and an “all eligible pages” mode. * Introduced `DOCS_E2E_PAGE_PATHS` (and updated base ref/base URL behavior) to control which pages are tested. * **Bug Fixes** * Automatically skips Playwright setup when no relevant pages are in scope; Playwright reporting now uploads only on failure. * **Documentation** * Updated the Docs E2E README with new run/CI behavior, troubleshooting notes, and commands to inspect the resolved page list. * **Tests** * Added a Docs-owned pages E2E suite; removed the Next.js quickstart E2E spec. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com> |
||
|
Miranda Limonczenko
|
b5cae478bc |
fix(docs) Add smoke test for local development without credentials (#48218)
Closes DOCS-1210 Closes DOCS-1209 #48226 needs to merge first for CI failure ## 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? Test coverage and several small bug fixes discovered during implementation. ## What is the current behavior? Nothing verified that `pnpm run dev:docs` keeps working without private credentials. We value this command working, especially for community contributors. However, this issue can go undetected by employees at Supabase since many of us have credentials in place. We do not want this to go a week before finding and fixing like in the previous instance. ## What is the new behavior? - **New Playwright test suite**: `e2e/docs/local-smoke/no-credentials.spec.ts` boots the docs dev server with zero GitHub App/Supabase secrets and checks 5 routes covering each known failure point. - **CI**: a new `local-dev-smoke` job in `docs-tests.yml` runs this suite with no credentials configured. ## Additional bugs resolved Setting up this test exposed other issues that are fixed in this PR: - **Troubleshooting.utils.ts crash** — Unguarded Supabase call pattern, crashing every troubleshooting article. Added the same guard as previous fixes. - **Missing manifest.json** — middleware.ts statically imports public/markdown/manifest.json, which is gitignored and only generated by a build step that's skipped in local dev. On a fresh checkout it doesn't exist, so middleware fails to compile and takes down every page. Fixed by committing a placeholder [] (real builds still regenerate the full file). - **Phantom @code-hike/mdx import** — apps/docs/app/layout.tsx imported @code-hike/mdx/styles.css, but only apps/www actually declares that dependency. Worked by accident whenever both apps were installed together; broke in CI's docs-only install. Turned out to be dead code (nothing in docs actually uses code-hike), so fixed by deleting the unused imports rather than adding the dependency. <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit ## Summary by CodeRabbit * **New Features** * Added a credential-free “local smoke” end-to-end test suite for key documentation routes. * **Bug Fixes** * Improved troubleshooting behavior when required external service credentials are missing. * Updated federated “wrappers” documentation pages to gracefully show a fallback message when external content can’t be fetched. * **Tests** * Added a dedicated local-smoke Playwright runner and enhanced CI path-based triggering and reporting (failure-focused artifacts). * **Chores** * Refined docs workflow path filters and adjusted docs markdown manifest/ignore rules for generated content. <!-- end of auto-generated comment: release notes by coderabbit.ai --> |
||
|
Miranda Limonczenko
|
9199aad57e |
feat(docs) Add scaffolding and CI/CD step for Docs Playwright (#48120)
Closes DOCS-1197 ## I have read the [CONTRIBUTING.md](https://github.com/supabase/supabase/blob/master/CONTRIBUTING.md) file. YES ## Problem We do not have any E2E testing established. ## Solution This PR creates an ultra-lean starting place for Docs Playwright: - A CI/CD step that skips on draft and relies on Preview for testing - One simple broken link check for one page The goal: - Playwright is implemented where we want it, with an architecture we want, with set-up steps we can build from The anti-goal of this PR: - We have meaningful tests running ## CI/CD steps <img width="1191" height="72" alt="Screenshot 2026-07-21 at 10 17 06 AM" src="https://github.com/user-attachments/assets/eeb2454c-d864-4574-a050-ce39bb3f083f" /> 1. Checkout a thin slice of the repo (`apps/docs`, `packages`, `patches`). 2. Wait for the Vercel **docs** preview for that commit SHA. 3. Use that preview URL as `PLAYWRIGHT_BASE_URL`. 4. Install Node deps and Chromium. 5. Run `pnpm run e2e:docs` (`--grep @quickstart`). 6. If anything fails, upload the HTML report + traces. Manual runs skip the Vercel wait and default to `https://supabase.com` (or whatever URL you enter), then run the full suite (`pnpm run e2e`). ## What the test checks Because this PR is scaffolding, it is doing something very basic: 1. Opens `/docs/guides/getting-started/quickstarts/nextjs` only if a connected file was edited in CI/CD step 2. Asserts the page loaded and the H1 is visible. 3. Collects docs-owned `/docs/**` links from `#sb-docs-guide-main-article`. 4. HTTP-checks each link (no full navigation) and soft-fails so every broken link is reported. Config keeps it cheap: Chromium only, 1 worker, 2 CI retries, failure screenshots/traces. ## Docs vs Studio/Dashboard The setup of Docs Playwright differs from Studio. | | Docs E2E | Studio E2E | |---|---|---| | Location |`e2e/docs/` | `e2e/studio/` | | What it tests | One published docs page + its links | Many Studio UI flows (tables, auth, storage, …) | | Where the app runs | Already-deployed **Vercel preview** | Built and started **on the runner** | | Backend needed | None | Local Supabase via Docker | | Path filtering | Native `on.pull_request.paths` (skip whole workflow) | `dorny/paths-filter` after checkout (workflow starts, heavy steps gated) | | Parallelism | 1 worker, no shards | Matrix of frameworks × 2 shards | | Retries | 2 in CI | 5 in CI | | Reports | HTML report on failure | Blob reports per shard → merge → PR comment | | Draft handling | Explicit draft skip | No draft skip today | | Manual broader run | Yes (`workflow_dispatch`) | No | The big conceptual difference: **Studio owns the environment** (build Studio, start Supabase, hit `localhost`). **Docs borrows Vercel’s preview** and only asks “does this page and its docs links work on the deployed site?” ## Docs architecture justification The docs architecture is deliberately lightweight because docs are **static, published content served by Vercel**, not an interactive app with a backend. That single fact justifies every difference: - **Borrow the Vercel preview instead of building on the runner.** The preview is already the exact artifact users will see, and Vercel builds it for free on every PR. Rebuilding docs on the runner would duplicate that work and risk testing something different from what ships. Studio, by contrast, needs a running app plus a local Supabase, so it *has* to own its environment. - **No backend.** Docs pages don't need a database or auth to render, so there's nothing to spin up. This is what keeps the job cheap enough to run per-PR. - **Native `paths` filtering.** Since the job is cheap and self-contained, an all-or-nothing skip at the workflow level is sufficient—no need for `dorny/paths-filter` to gate expensive setup steps mid-run like Studio does. - **Low parallelism and modest retries.** One page and its links is a tiny surface, so 1 worker is plenty and there's no sharding to coordinate. Retries exist only to absorb transient network flakiness against a live URL, hence 2 rather than Studio's 5 (which also cushions a heavier, stateful environment). - **Non-blocking + draft skip + manual dispatch.** As initial scaffolding checking link health on a deployed site, it should inform rather than gate merges, avoid burning minutes on drafts, and still be runnable on demand against production. In short: **Studio owns its environment because it must; docs borrows Vercel's preview because it can.** The scope is intentionally minimal today. ## Testing 1. Break a docs-owned link in the Next.js quickstart. 1. Follow README instructions to set up and run e2e docs test. 1. Confirm the suite fails. 1. Restore the broken link and re-run. 1. Confirm the suite **passes** (`1 passed`). <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit ## Summary - **New Features** - Added a GitHub Actions workflow to run Playwright docs end-to-end tests on PRs and via manual dispatch (with optional base URL), including docs-preview waiting and concurrency cancellation. - **Documentation** - Added `e2e/docs` README with setup, how to run the suite (including UI/debug and single-spec), and how base URL selection works. - **Tests** - Added a quickstarts E2E spec that validates the page and soft-checks docs-owned links resolve. - **Chores** - Added shared Playwright configuration/package scripts and an `e2e/docs` `.gitignore` for test outputs. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> |