Files
supabase/apps/studio/AGENTS.md
Alaister Young f125126aec chore: make agent instructions agent-agnostic (#49941)
Makes the repo's AI-agent setup tool-agnostic: instructions live in
`AGENTS.md` files, skills live in `.agents/skills/`, and Claude Code,
Codex, Cursor, and Copilot all read the same sources. Also sweeps the
skills for stale and duplicated content while everything was being
moved.

**Changed:**
- Every `CLAUDE.md` (root, `apps/studio`, `apps/docs`, `apps/kb`) is now
a one-line `@AGENTS.md` import; the content moved verbatim into an
`AGENTS.md` beside it. The root one moved from `.claude/CLAUDE.md` to
the repo root for consistency.
- All skills now live in `.agents/skills/`; `.claude/skills` is a single
symlink to it (replacing the old mix of real dirs and per-skill
symlinks). Path references in `.coderabbit.yaml`, code comments, and
docs updated to match.
- `.github/copilot-instructions.md` keeps only the review policy and
points at `AGENTS.md` + `.agents/skills/`. Copilot code review reads
those natively now, so the per-topic
`.github/instructions/*.instructions.md` files were duplicates of the
skills.
- Stale skill content fixed: `studio-queries` imported a toast library
Studio doesn't use, `telemetry-standards` and `studio-testing` used
import paths that don't resolve, `safe-sql-execution` cited a boundary
test that doesn't exist, the ask-the-docs references described an
`AiPrompt` mechanism that was replaced by the ID-keyed registry, plus a
handful of wrong paths, a self-contradicting `waitForTimeout` rule, an
invalid Playwright signature, and a ConfigCat flag described as PostHog.
- `studio-error-handling` now explains when to use `AlertError` (the
default) vs `ErrorMatcher`.

**Added:**
- `apps/docs/AGENTS.md` (docs test requirements, from the old Cursor
rule)
- `studio-shortcuts` skill (from the old Copilot instruction file,
verified against the current registry)
- `ask-the-docs/reference/graphql-endpoint.md` and
`search-embeddings.md` (from the old Cursor rules, with the missing
resolver/registration/codegen steps filled in)
- Feature-flag measurement section in `telemetry-standards`

**Removed:**
- `.cursor/` (rules folded in as above; skill symlinks no longer needed)
and `.cursorignore`
- `.github/instructions/` (8 files)
- `vercel-composition-patterns/AGENTS.md` – a 946-line verbatim
concatenation of its own `rules/` directory, and a nested `AGENTS.md`
that agents could auto-load as repo instructions
- `edit-the-docs/reference/structure-and-flow.md` – word-for-word copy
of the skill's own Phase 2 text

## To test

- `readlink .claude/skills` → `../.agents/skills`, and `ls
.claude/skills/copywriting/SKILL.md` resolves
- Open a Claude Code session at the repo root and in `apps/studio` – the
imported `AGENTS.md` content should load as before
- `git diff master --stat -M` shows the skill moves as 100% renames
(content unchanged except the listed fixes)
- Spot-check a fixed claim, e.g. `import { toast } from 'sonner'` in
`studio-queries`, or the `logs.all` ESLint rule cited in
`clickhouse-logs-queries/references/codebase-integration.md`

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

- **Documentation**
- Expanded guidance for documentation workflows, GraphQL resources,
search, ClickHouse logs, React forms, Studio testing, shortcuts,
telemetry, accessibility, copywriting, and composition patterns.
- Clarified local testing, linting, build workflows, error handling, and
AI coding agent usage.
- Added contributor guidance for the knowledge base, documentation, and
Studio areas.

- **Chores**
  - Consolidated agent instructions and skill references.
- Removed obsolete editor-specific guidance, duplicate links, and
superseded documentation.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

---------

Co-authored-by: Alaister Young <10985857+alaister@users.noreply.github.com>
2026-09-03 21:58:29 +08:00

11 KiB
Raw Permalink Blame History

Supabase Studio

Next.js pages router + TanStack Start (mid-migration, see below), React 19. Dev server: pnpm dev:studiohttp://localhost:8082.

Skills — load before working

Load the skills matching the task; stack them when a task spans areas:

Task Additional skills
Query/mutation hooks, query keys (data/**) studio-queries
UI: pages, forms, tables, charts, sheets, empty states studio-ui-patterns
Form logic: react-hook-form fields, watch/formState, reset, number inputs react-hook-form
Displaying API errors studio-error-handling
Tests (deciding, writing, reviewing) studio-testing, then studio-mock-api-tests (component/MSW) or studio-e2e-tests (Playwright)
PostHog event tracking telemetry-standards
SQL against user databases safe-sql-execution
Logs Explorer SQL, data/logs clickhouse-logs-queries
Component API design, boolean-prop refactors vercel-composition-patterns
Keyboard shortcuts, search/filter inputs studio-shortcuts
User-facing copy copywriting

TanStack Start migration

Studio is migrating from the Next.js pages router (pages/**) to TanStack Start (routes/**). Both runtimes ship side-by-side; the STUDIO_FRAMEWORK env var selects which one pnpm dev/build runs (default: next, resolved in scripts/dispatch.js). Full route map and strategy: TANSTACK_MIGRATION.md.

  • Never delete a page file. Most routes/** files are thin wrappers re-exporting the default export of their pages/** counterpart, so the Next file is load-bearing for both runtimes until the final cleanup pass.
  • Pure page-body edits propagate to the route automatically. Mirror a change by hand into the corresponding routes/** file only when it touches what the route duplicates: getLayout/layout wrapping, page titles or other staticData (incl. skip*Layout flags), withAuth, or redirect paths.
  • A new page under pages/** needs a matching route under routes/** plus a checklist entry in TANSTACK_MIGRATION.md.
  • New code uses native TanStack APIs — no next/router or next/link. The compat/next/ shims exist only for legacy re-exported pages.
  • routeTree.gen.ts is generated by the Vite plugin — never hand-edit.

Orientation

  • Data layer — all platform API calls go through data/fetchers.ts (openapi-fetch, typed by the generated api-types package) with handleError; never raw fetch. One folder per resource in data/, most with a keys.ts query-key factory.
  • State — valtio for global state (state/), nuqs for URL state, react-hook-form + zod for forms.
  • Platform vs self-hostedIS_PLATFORM gates platform-only behavior; withAuth is a no-op when self-hosted.
  • TelemetryuseTrack() from lib/telemetry/track; event types live in packages/common/telemetry-constants.ts.
  • Tests — default to including relevant tests with any change: a couple of unit tests for extracted logic, component tests for UI behavior, E2E only when the scope demands it (studio-testing has the decision tree). Not every PR needs them, but "no tests" should be a considered choice, not the default. Tooling: vitest + MSW; component tests use customRender + addAPIMock from tests/lib/; unhandled network requests fail tests. Don't vi.mock('@/data/...').
  • Shortcuts — use the registry in state/shortcuts/ and components/ui/Shortcut*.tsx; keep G then … chords for navigation; no one-off keyboard listeners (studio-shortcuts has the full pattern, including the search-input Escape handler).
  • Scoped PAT catalogpackages/shared-data/scoped-access-token-permissions.ts feeds Studio and the generated Personal Access Tokens guide. After changing it, run make -C apps/docs/spec generate.partials.access-control; Docs Tests rejects stale tables.
  • Reuse first — before writing a new hook or helper, search for an existing one (hooks/, lib/, packages/common, packages/ui-patterns). If you do need a new one, make it as reusable as possible: general naming, no page-specific coupling, placed where other callers can find it.
  • Co-locate sub-components with their parent; avoid barrel re-export files.

Code style

Older Studio code predates some of these conventions. For new or modified code, follow them rather than mirroring nearby legacy patterns:

  • Booleans read as is/has/can/should. Derive them from existing state (const isFormValid = name.length > 0 && email.includes('@')) — mirroring a derivable value into useState synced by useEffect is a bug pattern. Give multi-condition logic a name (const canShowAddButton = !isSchemaLocked && canUpdateColumns && …) instead of inlining the chain in JSX.

  • Ternaries: one is fine for a binary choice; never nest them. Anything bigger flattens — early returns in statement position, sibling && blocks in JSX.

  • Fetch states render with early returns at the top level, or a flat && chain with mutually exclusive guards inline — never a nested ternary:

    // Top level: early return per state
    if (isLoading) return <GenericSkeletonLoader />
    if (isError) return <AlertError error={error} subject="Failed to retrieve data" />
    if (isSuccess && data.length === 0) return <EmptyState />
    return <DataDisplay data={data} />
    
    // Inline: flat `&&` blocks, mutually exclusive guards
    <div>
      {isLoading && <ShimmeringLoader />}
      {isError && <AlertError error={error} />}
      {isSuccess && data.length === 0 && <EmptyState />}
      {isSuccess && data.length > 0 && <DataDisplay data={data} />}
    </div>
    
  • useEffect is for synchronizing with external systems (subscriptions, DOM, timers) — not for deriving data (compute it in render), reacting to user actions (do it in the handler), or fetching (React Query). Older code uses effects for all of these; don't copy it.

  • State stays as local as possible — lift it only when it's actually shared. Related form fields belong in a single react-hook-form + zod form, not parallel useState calls.

  • Component size: split at ~200300 lines — or sooner when a component grows multiple distinct UI sections, tangled conditional rendering, or clusters of unrelated useState. Extract repeated JSX into small components, non-trivial pure logic into .utils.ts functions (which get unit tests), and reusable stateful logic into custom hooks.

  • Memoization is not the default: useMemo/useCallback only for measured expense or referential stability a memoized child depends on.

  • TypeScript: avoid as casts — where external data enters, parse it with zod (schema.parse/safeParse) instead. Model multi-state values as discriminated unions ({ status: 'success'; data: T } | { status: 'error'; error: Error }) rather than independent boolean flags.

  • Naming: prop callbacks are onX, internal handlers are handleX. Custom hooks return objects, not tuples.

  • Refactoring: when you move or extract code into a new module, update every importer to point at the new location directly — do not leave a re-export shim in the old file "for backward compatibility." It's a one-line import change per consumer, and keeping shims around makes the codebase messy and the true source of a symbol ambiguous.

Defaults that differ here

  • ESLint warnings are ratcheted in CI: the per-rule occurrence count must not increase, so a new any, unresolved exhaustive-deps warning, or default export fails the build even though it's "only a warning". Check locally with pnpm --filter studio run lint:ratchet.
  • Dead files and deps are gated in CI by knip (pnpm knip --workspace apps/studio locally). Framework-convention files nothing imports (routes, Vercel functions, TanStack Start files) belong in knip.jsonc under workspaces["apps/studio"].entry, not ignore — an ignored file's imports aren't traced, so anything only it uses gets reported as dead. There's no inline knip-ignore comment; the only per-file opt-out is config. In a PR stack, a file whose first consumer lands in a later PR fails the gate on the earlier one — either add it in the same PR as its first use, or add it to workspaces["apps/studio"].ignore with a used from #NNN comment and remove it in that PR.
  • Clipboard: copyToClipboard from 'ui', and never await anything before calling it (Safari requires the write inside the user gesture; lint-enforced) — pass a Promise as the argument instead.
  • useParams() comes from 'common', not next/navigation — it camelCases keys and returns string | undefined.
  • Permissions: useAsyncCheckPermissions from hooks/misc/useCheckPermissions (returns can: true when self-hosted).
  • Gating: useIsFeatureEnabled for product features, useFlag from 'common' for feature flags — two different systems.
  • Dates: dayjs (plugins pre-loaded at both entries, pages/_app.tsx and routes/__root.tsx), not date-fns. Toasts: toast from 'sonner'.
  • Import split: 'ui' = primitives, 'ui-patterns' = composed patterns (ConfirmationModal, …), @ui/* = alias into packages/ui/src. Icons come from lucide-react.
  • New tables use @tanstack/react-table; react-data-grid is banned for new code.
  • Ad-hoc SQL against the user's database goes through executeSql / useExecuteSqlMutation (data/sql/execute-sql-mutation).
  • Confirmations: ConfirmationModal / TextConfirmModal from ui-patterns, never window.confirm. Disabled buttons needing an explanation use ButtonTooltip; inline warnings use Admonition.