chore: adopt the shared magic tooling stack

[GSTJ]

Moves pegada onto the shared GSTJ tooling from GSTJ/magic: the same oxlint preset, oxfmt config, tsconfig bases, CI job and Renovate preset every other repo in the set now uses.

Stack

	before	after
oxlint	1.64.0, hand-written .oxlintrc.json	1.75.0 + magic-oxlint-config (base / expo / next)
oxfmt	0.49.0, .oxfmtrc.json	0.60.0 + magic-oxfmt-config
tsconfig	local tools/tsconfig package	magic-tsconfig (the two files were byte-identical)
pnpm	9.6.0	11.17.0
Node	20.x	22.x
CI checkup	3-way matrix, inline steps	GSTJ/magic/.github/workflows/ci.yml@main
Renovate	local subset	github>GSTJ/magic

There was no ESLint or Prettier here to remove — pegada was already on oxlint. What changed is where the config comes from.

Three configs, not one. oxlint resolves the nearest config walking up from each file, and a nested config replaces the root rather than merging with it. apps/mobile gets expo, apps/nextjs gets next, everything else gets base from the root. All three restate ignorePatterns, because extends drops them.

pnpm 11 moved the settings. The pnpm key in package.json is no longer read, so overrides, patchedDependencies and supportedArchitectures are in pnpm-workspace.yaml now, along with the two .npmrc settings. It also hard-fails on unapproved dependency build scripts (allowBuilds) and refuses packages published in the last 24h (minimumReleaseAgeExclude, needed because the magic-* releases are same-day). Node had to go with it: pnpm 11.17 declares engines.node >= 22.13 and every workflow pinned 20.x.

Renames

96 files renamed, 357 import specifiers rewritten by magic-kebab, run with --tsconfig apps/mobile/tsconfig.json so the @/* aliases resolved. 0 conflicts, 0 skipped. Rename-only commit, so git log --follow still works.

Two things the codemod reports but deliberately does not touch, fixed by hand in the same commit because the tree is broken without them:

• workspace subpath imports (@pegada/api/services/PaymentService), which go through a package exports map rather than a relative path;
• the Expo config-plugin paths in app.config.ts, which are plain strings in the plugins array. A miss there fails only on a Linux/EAS build, since APFS is case-insensitive.

Violations

739 errors on first contact, now 0. The ones worth reading:

• 63 inline styles moved into the co-located styles.ts as styled components. Six form screens were each spelling out style={{ flex: 1 }}, so Fill / Row / KeyboardScreen now live in @/components/layout. Style props that cannot be a component — contentContainerStyle, the react-navigation header slots — became StyleSheet.create entries beside them.
• packages/api had an import cycle between the services, match and swipe modules. transformDistanceBetweenUserAndDog moved to shared/dog-distance.ts, where its three callers reach it without going through SwipeService.
• packages/database/__mocks__ is now fixtures/. Nothing in it was a jest mock; it is seed data three suites import directly, which is exactly what jest/no-mocks-import was reporting.
• A real bug in packages/database/seed.ts: 100 fake users were created inside a forEach passed to Promise.all, so nothing awaited them and the seed could return while they were still being written.
• 23 async functions with nothing to await lost the keyword. Two changed behaviour and were fixed rather than reverted — customNotificationHandler is synchronous, so the .catch(sendError) at its call sites never saw the Invalid notification url it throws. Those are try/catch now.
• Error classes: LikeLimitReached → LikeLimitReachedError, and every class sets name, because instanceof does not survive tRPC's serialization boundary but the name does.
• env.ts in packages/database and the queue-driver vars in the API config module, so process.env is read in exactly one place per package.
• jest.clearAllMocks() deleted from three suites in favour of clearMocks: true in both runners.
• The last .jsx file is now .tsx.

Every eslint-disable became an oxlint-disable carrying a reason, and the 14 that no longer suppressed anything are gone.

oxlint --fix needed supervision

Four autofixes had to be undone or repaired. Worth knowing before the next repo:

1. unicorn/prefer-string-raw broke the Next build. It rewrote middleware.ts's matcher to a String.raw template literal; Next statically analyses that export and fails with Invalid segment configuration export detected. Nothing else catches it — lint, typecheck and tests were all green.
2. unicorn/explicit-length-check corrupted logic. data.size ? data.size : null became data.size > 0 ? … on a field whose type is "SMALL" | "MEDIUM" | "LARGE". The rule assumes any .size is a collection.
3. unicorn/catch-error-name renamed a shorthand property key. .catch((cause) => … { cause }) became .catch((error) => … { error }), changing the object's key; in another file it renamed a binding onto a const error already declared in the same block.
4. unicorn/no-array-reverse targets ES2023. .toReversed() on lib: ES2022 and Hermes; both call sites already reversed a copy, so the rule is off for the mobile app — the same call the preset makes for unicorn/no-array-sort.

Local overrides

All repo-specific, each carrying its reasoning in the config file: the gesture-handler scrollable import ban, expo-* and ./styles namespace imports, the two env boundary modules, the API's static-method service classes, the single-module error taxonomy, sequential seeding, unicorn/no-array-reverse, react/react-compiler (it only reports bailouts on reanimated, gesture-handler and react-hook-form) and no-warning-comments (this repo keeps TODOs as durable notes; the alternatives were deleting them or rewording them to dodge the keyword).

Verified

From a clean node_modules: pnpm install --frozen-lockfile, pnpm dedupe --check, lint 0 diagnostics, format --check clean, typecheck across all five packages, 73 tests green against a live Postgres, and next build producing all 9 pages. The release, e2e and deploy workflows are untouched apart from the Node bump.

https://claude.ai/code/session_01Fe92vvdc4R3F4BDBFoLcw1

[GSTJ]

Updated to the magic 1.1.0 round

magic-oxlint-config, magic-oxfmt-config, magic-tsconfig and magic-codemods all move 1.0.0 → 1.1.0. magic-oxlint-plugin isn't consumed here, so 1.0.1 doesn't apply.

Four commits: the bump, the tsconfig fallout, the lint-config cleanup, the interface → type conversion.

What the bump forced

interface → type, 104 diagnostics. typescript/consistent-type-definitions was already "error" in 1.0.0 but carried oxlint's default option, which enforces the opposite direction. 1.1.0 sets ["error", "type"], so every interface in the repo was suddenly an error. --fix did the mechanical part across 61 files.

Five of them are back to interface, each with a disable and a reason:

• CustomTypeOptions in packages/shared/i18n/types/types.ts augments i18next's module. A type alias there is TS2300: Duplicate identifier, not a style choice — typecheck caught it.

• VisitingCardProps, MatchActionBarProps, FeedbackProps, EmailInputProps extend a base type carrying a string index signature (reanimated's Animated.View props, react-native-svg's SvgProps, a styled TextInput). interface X extends Y keeps each declared member at its declared type; { … } & Y intersects every one of them with the index signature's any and widens them silently. startImageIndex?: number became any.

  tsc only caught one of the four, and indirectly: useState(startImageIndex) lost its contextual type and the setCurrentImage((index) => index - 1) callback tripped TS7006. The other three were found by asserting 0 extends 1 & T on every converted property — no error, no test failure, just three props that had stopped being typed. Worth knowing before running this fixer on another repo.

tsBuildInfoFile is now a hard error. magic-tsconfig 1.1.0 drops incremental from base.json, and tsBuildInfoFile without it is TS5069. All five package tsconfigs had it; typecheck failed on the bump alone until they came out.

Knock-ons: turbo's typecheck task declared that same .tsbuildinfo as its output, so it was about to start warning "no output files found" on every run — the exact thing this repo already fixed once for test. And next build writes suggested compiler options back into tsconfig.json when the key is absent, so it re-added "incremental": true and reformatted the file on every build, leaving a dirty tree and a failing oxfmt --check. apps/nextjs/tsconfig.json now sets "incremental": false explicitly, which stops that.

Redundant local config removed. The mobile config's unicorn/no-array-reverse: "off" — the preset ships it off now, same ES2023/Hermes reasoning. And middleware.ts's oxlint-disable for unicorn/prefer-string-raw, which the next preset's App Router override covers; --report-unused-disable-directives was reporting it.

On the ignorePatterns re-listing

I checked whether the manual ignorePatterns: [...base.ignorePatterns] lines could come out. They can't — extends still drops them on oxlint 1.75.0 with the 1.1.0 presets. Verified by linting a file under a generated/ directory (in the preset's ignore list, not in .gitignore) with and without the manual re-listing: without it, the file is linted.

Rather than keep three copies of a line that's load-bearing and easy to get wrong, the two configs with repo-specific rules now use extendConfig, which flattens, and apps/nextjs — which had nothing left but the boilerplate — is the one-line re-export the magic README asks for. Resolved diagnostics across the whole repo are byte-identical before and after the reshape.

One thing to be careful with: --print-config is not a reliable way to check this. Under extends it also omits rule options, categories and jsPlugins that do in fact apply at lint time — typescript/consistent-type-definitions prints as bare "deny" with no "type" option while enforcing it, and safe-jsx prints as absent while firing. Only ignorePatterns is genuinely dropped. Behavioural probes, not printed config.

Verification

Clean node_modules, pnpm install --frozen-lockfile → lint 0 diagnostics, oxfmt --check clean, typecheck 5/5, tests 73/73 (20 mobile + 53 api), pnpm dedupe --check clean. magic-kebab --dry-run --strict exits 0 with nothing to rename, and 1.1.0's workspace-walking discovery now finds both app tsconfigs instead of none.

pnpm run build still can't run locally — it needs .env.dev, which is gitignored, and turbo's strict env mode drops what dotenv sets for anything but test. Same failure on the branch's previous head, so it's not from this change. next build run directly with .env.test loaded succeeds, middleware included.

[GSTJ]

Moved onto the 1.2.0 round: magic-oxlint-config, magic-oxfmt-config and magic-tsconfig. magic-codemods stays at 1.1.0.

One workaround came out. apps/nextjs/tsconfig.json had incremental: false in it, added in this PR to stop next build writing the option in itself and reformatting the file. 1.2.0 puts incremental back in nextjs.json, so the override is redundant. Checked rather than assumed — a full next build now leaves tsconfig.json byte-identical and drops apps/nextjs/tsconfig.tsbuildinfo, which .gitignore already covers. The turbo.json comment about typecheck writing nothing was updated to match.

Nothing to convert for the extends deprecation. All three oxlint configs already use extendConfig() (root, apps/mobile) or the one-line re-export (apps/nextjs), from the 1.1.0 round.

No withoutIgnorePatterns() either. CHANGELOG.md here is generated by .github/scripts/changelog.py, and nothing calls oxfmt on it, so the default ignore is the right behaviour.

pnpm quarantine. Both versions were listed in minimumReleaseAgeExclude for the one install that rewrote the lockfile, then the 1.1.0 lines removed — pnpm install --frozen-lockfile is clean. magic-codemods@1.1.0 keeps its entry. The lockfile diff is 29 lines and all of them are the three packages.

Checks

check	result
install --frozen-lockfile	clean
lint	0 diagnostics, 363 files
format --check	clean, 464 files
typecheck	5/5
test	79 tests, 10 suites
build	fails locally, see below

pnpm build fails at @pegada/nextjs, on Invalid environment variables collecting page data for the queue routes — JWT_SECRET, the AWS keys, the PostHog keys and so on. It is not this bump: the same failure reproduces on the branch tip with these commits stashed, and turbo's strict env mode is what keeps .env.test from reaching the build task (only test declares passThroughEnv). Local-only, no .env.dev on this machine. Everything up to page-data collection compiles.