Run the tests

No mocks of the code: real browsers, the real pages, a real vault library against a mock of the vault API, and the sgit command line. The harness swaps one thing only: requests to this host are answered from the working tree.

Set up

git clone https://github.com/SGit-AI/SGit-AI__Website__Newsroom__API && cd SGit-AI__Website__Newsroom__API
npm install --no-save --no-audit --no-fund playwright@1.56.1
npx playwright install --with-deps chromium        # and webkit, firefox, as you need
python3 -m pip install 'sgit-ai==0.21.0'           # the vault test's CLI round trip, the notes test's lane
git clone https://github.com/SGit-AI/SGit-AI__Website__Newsroom .newsroom   # the pages the tests run against

tests/newsroom.json names the commit CI checks the newsroom out at. NEWSROOM=<path> points the tests at any checkout, including your own site once it imports the code.

Run

node tests/test_meter.mjs                       # Chromium, the default
BROWSER=firefox node tests/test_lists.mjs
BROWSER=webkit  node tests/test_search.mjs
NEWSROOM=/path/to/your/site node tests/test_meter.mjs

A test prints ok or FAIL per check and exits 1 on any failure. It prints SKIP and exits 0 when it cannot run (no browser, no newsroom checkout, no sgit); CI treats a SKIP as a failure, so a missing install cannot pass quietly.

What each file covers

Test	Covers	Engines
`test_styles`	the served stylesheets write no colour outside a theme token (the newsroom's rule, held here since the files moved)	Node
`test_site`	this website: the five themes' tokens, no literal colour in its stylesheet, the picker and its memory, 4.5:1 in every theme, the bar at ten widths	all three
`test_meter`	the meter in a browser: depth, rating, decline, reload, below zero, the return page, personas, picks, export and restore, the share form, the persona links, phone width	all three
`test_lists`	the reading-list parser in Node and the two pages in a browser: malformed input, limits, the make page, copy, phone width	all three
`test_review`	the review's findings, pinned as the behaviour is today (R1 to R11, R14, R15, R17)	all three; R2 is held in Chromium
`test_wire`	the wire pages reading the real vault: edition, lenses, sources, documents, what left the browser	all three (needs the vault host)
`test_graphs`	the wire's graphs: the lenses' edges and the picture drawn from the real vault	all three (needs the vault host)
`test_sync`	the activity vault's logic as pure functions, two devices' states	all three (Node)
`test_search`	the index and the ranking in Node, the page in a browser	all three
`test_vault`	the vendored library against the mock, the newsroom's layer, the sgit CLI round trip	Node + CLI
`test_passkey`	the passkey lab with a virtual authenticator that has PRF	Chromium (CDP)
`test_pane`	the pane: the account carried between three devices through the mock	Chromium (CDP)
`test_notes`	the feedback loop end to end: notes, the lane, the CLI reading it, a second device	Chromium (CDP)
`test_inbox`	the pane's inbox and send: a link held, a QR code decoded, a second device	Chromium (CDP)
`test_ask`	ask: passages and messages in Node; the reader's key and the faked built-in model in a browser	Chromium

Chromium's clipboard permissions have no equivalent in Firefox and WebKit, so the three clipboard reads run in Chromium only; the rest of each test runs everywhere.

In CI

.github/workflows/ci.yml runs the syntax check, the manifest check and the reproducible build, then the tests in a matrix of the three engines against the pinned newsroom commit, on every push and pull request; about four minutes. The brief's target is two; caching the engines is the first thing to measure.