Files
it-tools/renovate.json
SebastianandClaude Opus 5 ec740dd063 ci: rework the test and lint toolchain — per-browser E2E, scoped jsdom, type-aware oxlint (#561)
* ci(e2e): one job per browser with three workers

Replace the 3-way shard matrix with a per-browser matrix. Each job now
runs one Playwright project with three workers instead of three jobs each
running a mixed slice of all browsers at one worker.

The old layout gave each of three runners a single worker, so each job
used one of its four cores; a browser failure also showed up as
"shard 2/3" with no indication of which engine broke.

- `workers: 1` -> `3` on CI. Measured on a 4-core box: 9.4 min at one
  worker vs 5.1 min at three, with the box ~95% busy. A fourth worker
  added nothing and left no core for the preview server.
- `fail-fast: false` so one engine failing still reports the other two.
- Install and cache only the browser a job actually needs, instead of all
  three in every job.
- Fix the cache key, which read `.dependencies.playwright` and always
  resolved to `null`; the package is `devDependencies["@playwright/test"]`,
  so the key never changed when Playwright was upgraded.
- `set_node_mem.sh` grows a `NODE_MAX_OLD_SPACE_MB` override, used to cap
  each worker at 4GB. Playwright workers inherit `NODE_OPTIONS`, so
  otherwise every worker claims a machine-sized heap ceiling.

The script is now POSIX. It is sourced from package.json scripts, which
run under dash, where `[[` does not exist -- every run was logging
"[[: not found" and silently falling through to the Linux branch, so the
macOS detection never worked when invoked that way.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* refactor(set_node_mem): capture memory sizes before the arithmetic

Addresses review feedback on the sourced heap-sizing script. Reads
closer to the original, which also assigned the raw value first.

Behaviour is unchanged: dash still produces --max-old-space-size=15051
by default and honours NODE_MAX_OLD_SPACE_MB.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* fix(e2e): give CI more timeout headroom now that workers share a runner

The webkit job failed on `markdown-diff`, which mounts a Monaco diff
editor -- the heaviest tool in the suite -- on the slowest engine:

  1 failed   [webkit] markdown-diff Has correct title
  1 flaky    [webkit] markdown-diff Diff editor is visible with two panes
  107 passed (9.3m)

Both are first-paint timing, not behaviour. The 15s expect timeout was
sized for one worker per runner; with three sharing four cores, first
paint of a lazy-loaded tool takes longer, and the failure survived two
retries.

Raise the CI budget to 30s for assertions. The test timeout has to move
with it -- it defaults to 30s, so a 30s expect timeout would trip the
test timeout first and report a timeout instead of the failing
assertion. Local runs keep the previous 15s/30s.

Timeouts are upper bounds, so passing tests are unaffected.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* fix(e2e): drop the inline timeout that caps markdown-diff below the CI budget

Raising the CI expect timeout to 30s fixed one of the two webkit
failures in markdown-diff but not the other:

  before: 1 failed + 1 flaky, 107 passed
  after:  1 failed,           108 passed

The surviving failure asserts the title with an inline
`{ timeout: 15000 }`, which overrides the global expect timeout, so it
kept the old budget while its neighbour picked up the new one.

The inline value only restated what the global used to be, so removing
it is a no-op locally (still 15s) and lets CI use its 30s budget.

The four inline overrides in mermaid-exporter are left alone: those
raise the timeout to 30s rather than capping it.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* perf(test): scope jsdom to the suites that need it

`test:unit` ran `vitest --environment jsdom`, which builds a jsdom for
every test file. Only 7 of 128 need one -- the other 121 are pure
functions and pass unchanged under the node environment.

Measured on a 4-core box, `pnpm test`, warm cache:

  before  89s wall, environment 128.79s, tests 4.70s
  after   ~56s wall, environment   7.37s, tests 5.30s

Environment setup was the single largest cost in the run and is now
~7s. The tests themselves were always ~5s.

Default to the node environment and let the 7 DOM-dependent suites opt
in with a `// @vitest-environment jsdom` docblock:

  src/ui/c-input-text                      (@vue/test-utils mount)
  src/ui/n-input-number-i18n               (@vue/test-utils mount)
  src/tools/certificate-key-parser         (webcrypto-liner-shim)
  src/tools/extract-text-from-html
  src/tools/markdown-preview
  src/tools/file-splitter
  src/tools/django-secret-generator

Same 128 files and 1035 tests pass before and after.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* test: drop jsdom from three more suites and fix two assertions

Down from 7 jsdom suites to 4. The three removed did not need a DOM;
they needed it by accident.

certificate-key-parser
  The service and its infos module each did a side-effect
  `import 'webcrypto-liner-shim'`. That shim polyfills WebCrypto for
  browsers without a secure context (see README: WebCrypto needs HTTPS),
  and it touches browser globals at import time, so pulling it into a
  node process fails before any test runs.

  Node has WebCrypto natively, so the polyfill only belongs on the
  browser path. Move the import to certificate-key-parser.vue, which is
  what pgp-keygen.vue and pgp-encryption.vue already do. It now also
  evaluates before sshpk rather than after it, and no longer loads
  eagerly through the tool's infos module.

  That let the test drop its jsdom workaround too: it was replacing
  global.TextEncoder/TextDecoder with the text-encoding-utf-8 polyfill,
  which node does not need. The `no-restricted-globals` disable existed
  only for those writes and is gone with them.

  Side effect: the run no longer emits the shim's asm.js linking
  warning, which was the only warning left in the unit suite.

extract-text-from-html
  `validateHtml` parsed the input with DOMParser, discarded the
  document, and validated with a regex instead. `parseFromString` with
  'text/html' recovers from bad markup rather than throwing, so the
  catch was unreachable and the whole block was dead -- and the only
  reason the file needed a DOM.

django-secret-generator
  Used `window.crypto`; `globalThis.crypto` is the same API in browsers
  and in node.

Two assertions did not test what they claimed:

  expect(getTextFromHtml('<p>..</p>')).toString();

`.toString()` on the assertion object asserts nothing -- this passed
even if the function returned null. Now checks the type.

  expect(getSecretKey()).toMatch(/[a-z0-9!@#$%^&*(-_=+)]{50,60}/);

`toMatch` searches rather than anchors, so any key of 50 chars or more
passed regardless of length. Split into an explicit length range and an
anchored charset check. The old class was also wrong: `(-_` reads as the
range `(` to `_`, which accepts digits and uppercase.

Also moved a valid-html case that sat in the invalid block, and fixed
the `invlid` / `django-secret-generato` typos.

128 files pass, 1035 tests -> 1036 (the django assertion is now two).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* chore(deps): drop the now-unused text-encoding-utf-8

The certificate-key-parser test was the only consumer: it replaced
global.TextEncoder/TextDecoder with this polyfill so the suite could run
under jsdom. That workaround went away when the suite moved to the node
environment, which provides both natively.

pnpm also prunes an orphaned transitive crypto-js@4.2.0, whose own
metadata marks it as no longer maintained. The lockfile change is 22
deletions and no additions.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* ci: enable oxlint type-aware linting, drop typecheck from build and CI

Adds oxlint-tsgolint and turns on `options.typeAware`, which brings 59
type-aware rules from typescript-eslint into `pnpm lint`. It found 59
issues; `--fix` resolved 3 (useless `= undefined` defaults).

The remaining 56 are semantic -- floating promises, base-to-string,
misused string spread -- spread over ~20 files, so they are staged as
`warn` rather than fixed blind in a CI change. Lint exits 0 and the
findings are visible in the output, ready to burn down by rule.

TypeScript stays on 6.0.3. The docs ask for TypeScript 7+, but that is
about tsgolint's own toolchain, not the project's: oxlint-tsgolint ships
a self-contained binary with no npm dependencies, and produces the same
56 findings against TypeScript 6.

Upgrading the project to TypeScript 7.0.2 was tried and reverted -- it
breaks the build:

  [@vue/compiler-sfc] No fs option provided to `compileScript` in
  non-Node environment. File system access is required for resolving
  imported types.

Isolated by rebuilding with the same scripts on 6.0.3, which passes. TS
7's native port also drops the `./lib/tsc` subpath export, so vue-tsc
cannot start at all -- neither 3.3.6 nor the current 3.3.10.

Script and workflow cleanup:

- `build` no longer runs `vue-tsc --noEmit` first; it just builds.
- The `ci` workflow drops its Type check step. It already runs `pnpm
  lint`, which now carries the type-aware rules. No other workflow had
  a typecheck step.
- `precommit` and the PR template drop `pnpm typecheck` for the same
  reason.

The `typecheck` script itself is kept, and so is vue-tsc. tsgolint
analyses no .vue files at all -- 0 of the 59 findings came from one --
so vue-tsc remains the only thing that type-checks single-file
components, and deleting it would drop that coverage with nothing
replacing it. It is simply no longer wired into the build or CI.

Renovate's TypeScript toolchain group gains oxlint-tsgolint, whose
version tracks TypeScript releases.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* refactor(oxlint): move config to oxlint.config.ts

Replaces .oxlintrc.json with a typed config built from oxlint's own
`defineConfig`, so rule names and options are checked by the editor
instead of being free-form JSON strings, and the staged type-aware rules
can carry an explanation next to them.

Verified as a faithful translation by diffing `oxlint --print-config`
between the two: identical resolved rule set (167 rules), no rule added,
dropped or re-levelled, and `options.typeAware` equal. Same 56 warnings
and 0 errors either way.

TypeScript configs are auto-discovered but marked experimental upstream
and require running under Node, which `pnpm lint` already does.

Also records why vue-tsc stays, since oxlint's vue plugin invites the
opposite conclusion: those 46 vue rules are style and correctness lint
rules, not a type checker. Planting

  const deliberateTypeError: number = getSecretKey();

in a .vue script block is reported by neither `--type-aware` nor
`--type-check` -- oxlint exits 0 -- while vue-tsc catches it as TS2322.
tsgolint only analyses .ts files.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* ci: restore typecheck and run lint + typecheck on every build path

oxc-project/oxc#15761 ("oxlint: better vue support") is the tracking
issue for Vue in oxlint. It is open, unimplemented, and scoped to
parsing template bindings so structural lint rules can work -- not to
type checking. Even once it ships it would not cover what vue-tsc does,
and the maintainer notes eslint-plugin-vue cannot be fully replicated
because Vue templates carry no JavaScript bindings for the linter.

So vue-tsc stays and goes back into CI. Confirmed the gap is real by
planting `const x: number = getSecretKey()` in a .vue script block:
oxlint exits 0 under both --type-aware and --type-check, vue-tsc reports
TS2322.

Every job that runs `pnpm build` directly now runs lint and typecheck
first:

  sharevb-ci                        + typecheck
  sharevb-docker-nightly-release    + typecheck
  sharevb-docker-realease-latest    + typecheck
  sharevb-docker-realease-baseurl   + typecheck
  sharevb-github-pages-publish      + lint, typecheck  (had none)
  sharevb-releases                  + lint, typecheck  (had none)

The two publish workflows were building and shipping without any check
at all. Left alone: the nightly github-release job, already gated by
`needs: ci` on the same commit, and the e2e workflow, whose three
browser jobs build on the same commit that the ci workflow is already
checking in parallel.

`precommit` and the PR template get typecheck back for the same reason.

oxfmt config moves to oxfmt.config.ts alongside oxlint's, using oxfmt's
own defineConfig. Same 832 files are flagged before and after, so the
translation is faithful.

Format checking is deliberately NOT wired into CI yet: `pnpm fmt:check`
currently fails on 832 of 1613 files (410 .vue, 359 .ts, 50 .md, 12
other). The scripts exist but have never been enforced, so the tree has
drifted. Enabling the gate needs a repo-wide reformat, which belongs in
its own commit rather than buried here.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* ci: report formatting drift without enforcing it

Adds `pnpm fmt:check` to the ci workflow with continue-on-error, so the
832 files that predate oxfmt being run in CI are visible on every PR
without blocking anything. Drop continue-on-error once a repo-wide
reformat lands.

Only the ci workflow gets it. It is the one that runs on pull_request,
so it is where a non-blocking signal is actually read; adding it to the
docker and release workflows would just add noise to logs nobody opens.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* refactor(test): pin the timezone in config instead of every script

`TZ=UTC` was repeated across four scripts, including once redundantly
(`test` set it and then called `test:unit`, which set it again).

It is genuinely load-bearing for the unit suite. Running vitest on
America/New_York fails 7 tests across 5 files -- date-time-converter,
days-calculator, crontab-generator and friends assert on formatted dates
and offsets. Moved to `test.env` in vite.config.ts, which is enough on
its own: the full suite passes with the host on America/New_York and on
Asia/Tokyo with no inline prefix.

That also fixes `pnpm coverage`, which never had the prefix and so would
fail those same 7 tests on any machine not set to UTC.

For e2e it was doing nothing. The browser timezone is already pinned by
`timezoneId: 'Europe/Paris'`, and the specs do no date maths on the node
side -- the one `Date.now` override runs inside `addInitScript`, so it
is browser-side. The timezone-sensitive specs pass on Asia/Tokyo without
it. Pinned in playwright.config.ts anyway so the runner stays
deterministic if a spec ever does compute a date in node.

Also switches `test` from `npm run` to `pnpm run`, since everything else
in the repo uses pnpm.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* chore: drop the `test` alias in favour of `test:unit`

`test` was a pass-through to `test:unit` and nothing else, so it only
added a level of indirection and a second name for the same thing --
confusing next to `test:e2e`, which is a different suite.

Repointed all five callers: the ci, nightly-release and two docker
release workflows, plus the README's "Run Unit Tests" snippet.

Note `pnpm test` no longer resolves, since npm's implicit test
lifecycle has nothing to fall back on now.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* refactor(e2e): fold test:e2e:dev into test:e2e

The two scripts were not two environments, they were two answers to
"who owns the web server":

  test:e2e      Playwright starts `pnpm preview` on :5050 and tests the
                production build
  test:e2e:dev  assumes `pnpm dev` is already up on :5173 and starts
                nothing

That is one variable, not a second workflow, so it does not need a
second script. The config now derives it: if E2E_BASE_URL points
somewhere other than the preview URL, you are supplying the server and
Playwright leaves the lifecycle alone. NO_WEB_SERVER is gone -- it only
ever restated what a custom base URL already implies.

Renamed BASE_URL to E2E_BASE_URL along the way. vite.config.ts reads
BASE_URL as the app's base path (`/it-tools/`, set by the Pages and
Docker builds), so one exported value meant two incompatible things and
would have pointed Playwright at a path instead of a URL.

Both paths verified:
  pnpm test:e2e                                  5/5, preview started
  E2E_BASE_URL=http://localhost:5173 pnpm ...    5/5, dev server reused

Documented both in the README, which had no e2e section at all -- the
script was the only place the dev flow was written down. Also notes that
a cold dev server compiles on demand and can blow the assertion timeouts
on a first run; that cost 2 failures until the routes were warmed.

Also fixes webServer.command, which shelled out to `npm run preview`.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* chore: delete set_node_mem.sh

The script raised the V8 heap to total RAM minus 1GB and was sourced by
build, test:unit and test:e2e. Nothing needs it any more.

The build does not come close to the heap. It completes on Node's
default (2096MB here) and still completes when capped at 512MB -- a
thirtieth of what the script was handing out:

  cap 15051MB (script)   92s, 93s
  default 2096MB         94s, 103s
  cap 1024MB             91s
  cap  512MB             95s

Unit tests are the same story: 35s and 37s with, 36s and 37s without.
No measurable difference anywhere, so the script was not buying speed
either.

The reason is that the build is not a JavaScript-heap workload. It peaks
around 6.5GB of system memory while running fine under a 512MB heap cap,
because rolldown does the work in Rust, outside V8. The README's "you
need around 16GB" dates from the old JS bundler; corrected to ~8GB with
the measured figure.

Removing it also retires two things it dragged along: the
NODE_MAX_OLD_SPACE_MB override added for Playwright, which existed only
because every worker inherited a machine-sized ceiling from this script,
and a POSIX bug where `[[` failed under dash on every single run.

Verified with NODE_OPTIONS unset, so at the 2096MB default CI will see:
build 87s, 129 files / 1072 unit tests, e2e 5/5.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

* fix(e2e): key the browser cache on the installed Playwright version

The key read the range from package.json, so it was `^1.62.0` rather
than `1.62.0`. Those agree today, but the range also matches 1.62.3 and
1.63.0: a lockfile bump inside the same range leaves the key unchanged,
so the run restores browsers for the old version and the cache is never
refreshed, because that key already exists.

Read node_modules/@playwright/test/package.json instead. pnpm/setup has
already installed dependencies by this step, so the resolved version is
available.

Raised by Copilot on the upstream PR.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YAwAJ6L3Mai8eSxXf5e6SA

---------

Co-authored-by: Claude <noreply@anthropic.com>
2026-08-18 21:55:58 +02:00

85 lines
2.1 KiB
JSON

{
"configMigration": true,
"branchConcurrentLimit": 5,
"prHourlyLimit": 25,
"recreateWhen": "never",
"prCreation": "immediate",
"timezone": "Europe/Paris",
"schedule": [
"* 0-3 * * *"
],
"automergeSchedule": [
"* 0-3 * * *"
],
"extends": [
"config:recommended",
"docker:pinDigests",
"helpers:pinGitHubActionDigests",
":pinDevDependencies"
],
"packageRules": [
{
"groupName": "TypeScript toolchain",
"matchPackageNames": ["typescript", "vue-tsc", "oxlint-tsgolint"],
"matchUpdateTypes": ["major", "minor", "patch"]
},
{
"groupName": "Vue ecosystem",
"matchPackageNames": [
"vue",
"@vue/compiler-sfc",
"@vue/runtime-dom",
"@vue/test-utils",
"@vue/tsconfig"
],
"matchUpdateTypes": ["major", "minor", "patch"]
},
{
"groupName": "Vite ecosystem",
"matchPackageNames": [
"vite",
"@vitejs/plugin-vue",
"@vitejs/plugin-vue-jsx",
"vite-plugin-pwa",
"vite-plugin-vue-markdown",
"vite-svg-loader"
],
"matchUpdateTypes": ["major", "minor", "patch"]
},
{
"groupName": "UnoCSS ecosystem",
"matchPackageNames": [
"unocss",
"unocss-preset-scrollbar"
],
"matchUpdateTypes": ["major", "minor", "patch"]
},
{
"groupName": "Unplugin ecosystem",
"matchPackageNames": [
"unplugin-auto-import",
"unplugin-icons",
"unplugin-vue-components"
],
"matchUpdateTypes": ["major", "minor", "patch"]
},
{
"groupName": "Vicons",
"matchPackageNames": ["@vicons/material", "@vicons/tabler"],
"matchUpdateTypes": ["major", "minor", "patch"]
},
{
"groupName": "VueUse",
"matchPackageNames": ["@vueuse/core", "@vueuse/head", "@vueuse/router"],
"matchUpdateTypes": ["major", "minor", "patch"]
},
{
"groupName": "Type packages with their dependencies",
"matchDepTypes": ["devDependencies"],
"groupSlug": "types",
"automerge": true,
"matchPackageNames": ["/^@types//"]
}
]
}