dualkit

Check any screen for Arabic and right-to-left bugs, and get a shareable Gulf-ready report.

npx [email protected] install chromium   # once
npx dualkit check https://example.com/ar

It opens the page in Chromium, runs 34 rules (bidi, layout, icons, type, content), prints a summary and writes dualkit-report.html with a score out of 100 and every issue marked on a screenshot.

| Option | Default | | | --- | --- | --- | | --out <file> | dualkit-report.html | Where the report goes | | --json | | Print the result as JSON | | --width, --height | 1280, 800 | Viewport, for example --width 390 --height 844 for a phone | | --min-score <n> | 0 | Exit 1 below this score, for CI | | --config <file> | dualkit.config.json | Ignore rules or regions | | --review | | Experimental. Also ask a vision model for problems the rules cannot see, see Review | | --review-model <provider:model> | gemini:gemini-flash-lite-latest | The model for --review |

{ "ignore": ["type/arabic-size"], "exclude": [".third-party-widget"] }

Every rule, why it matters and how to fix it: RULES.md. Exit codes: 0 done, 1 below --min-score, 2 a problem with the input, the browser or the page (including a page that answers with an error status such as 404).

Review

GEMINI_API_KEY=your-key npx dualkit check https://example.com/ar --review

--review is experimental. It has not been measured against hand labels yet, so treat its findings as suggestions. It adds a second pass after the rules. It sends screenshots of the page to a vision model and asks about problems a rule cannot measure, such as a date that reads in the wrong order or an icon that points the wrong way. Each problem it reports is then checked a second time on a close-up crop, and only the ones the model confirms firmly are kept.

By default it uses gemini-flash-lite-latest from Google, which has a free tier. Pick another model with --review-model provider:model, where the provider is gemini or openrouter, for example --review-model openrouter:google/gemma-3-27b-it:free.

The keys are read from your environment: GEMINI_API_KEY for Gemini and OPENROUTER_API_KEY for OpenRouter. The run uses your own key and your bill. The $3 a month cap on model spend applies only to Dualkit's own measurement runs, not to yours. The summary and the --json output say how much it spent, but Gemini reports no cost, so spend on a billing-enabled Gemini key is not counted. Without a key, the check runs as normal and one line says that review was skipped.

Review findings are warnings. They are marked review, they carry a confidence from 0 to 1, and they point at a source line when you check a local file. They never change the score or the exit code, and they never replace or remove a finding from the rules.

The model can be wrong. It can miss a real problem, and it can report one that is not there, so treat each finding as a prompt to look, not a verdict. Pages taller than four screens are only reviewed down to the fourth, and the run says so.

Lint your source

npx dualkit@beta lint .

dualkit lint [paths...] scans CSS, SCSS, HTML, Vue, Svelte, JavaScript and TypeScript against the 8 source rules that run by default, without launching a browser. Paths default to the current folder, and a folder walk skips what your .gitignore skips.

physical.css:2:3  · source/physical-css  `margin-left: 8px` is tied to one side. Use `margin-inline-start`.

0 errors, 0 warnings, 1 note in 2 files

| Option | Default | | | --- | --- | --- | | --json | | Print the result as JSON | | --fail-on <level> | error | Exit 1 on findings at or above error, warning or note | | --config <file> | dualkit.config.json | The same ignore list; it accepts lint rule ids too |

Exit codes: 0 clean, 1 a finding at or above --fail-on, 2 a problem with the input. The lint rules live in @dualkit/lint, which is published before the CLI.

In CI, uses: dualkit/dualkit/action@main with mode: lint runs it on every pull request: action/README.md.

Convert an app

npx dualkit@beta convert --mode bilingual --start "npm run dev" --port 3000

dualkit convert turns an English Next.js App Router or Vite React app into an Arabic one on a new git branch, dualkit/bilingual or dualkit/arabic. It needs a clean working tree, never pushes and makes no model calls. bilingual keeps English exactly as it is and adds Arabic with a switch. arabic makes the whole app Arabic and right to left, in place.

Branch: dualkit/bilingual
22 files changed, 52 edits, 10 strings to translate, 3 places left for you.

Fill the Arabic in dualkit-strings.json, or ask your agent to use the dualkit-convert skill. Then run:

npx dualkit@beta convert --finish

Finish compares each English page with its screenshot from before the conversion, runs the check rules on each Arabic page and writes dualkit-convert.md.

| Option | Default | | | --- | --- | --- | | --mode <mode> | | bilingual or arabic. Asked for on a terminal when left out | | --finish | | Run phase 2 and write the report | | --force | | With --finish, use English for strings with no Arabic yet | | --start <cmd> | | The command that starts your app, saved under convert in the config file | | --port <n> | | The port it listens on, saved with the start command | | --no-install | | Do not install next-intl or react-i18next | | --no-baseline | | Skip the English screenshots, so English is not proven unchanged. Arabic mode never takes them | | --config <file> | dualkit.config.json | Where the convert block is read and the start command saved. It must be inside the app folder. The block also takes test, routes, params and storageState |

Exit codes: 0 phase 1 done or finish passed, 1 finish failed, 2 refused or a problem with the input, 130 interrupted. What it converts, what it flags and the config block in full: @dualkit/convert.