verify

name	verify
description	Build InputStash, load it into a disposable Chrome, and verify capture/storage/popup behavior end-to-end, including input-lag measurements with a large stash.

Verify InputStash end-to-end

The extension's surface is a real browser: a content script capturing typing, a background service worker writing storage, and the popup UI. Verify by driving all three, not by unit-calling modules.

Build

pnpm build            # .output/chrome-mv3
pnpm build:firefox    # .output/firefox-mv2 (build-only check; no FF harness here)

Harness

harness/bench.mjs (in this skill directory) drives everything with puppeteer-core. One-time setup inside harness/:

npm init -y && npm i puppeteer-core@24 --no-audit --no-fund
npx @puppeteer/browsers install chrome@stable --path ./browsers
# then update CHROME in bench.mjs to the printed binary path

Gotchas learned the hard way:

• Branded Google Chrome ignores --load-extension (removed in Chrome 137). Use the Chrome for Testing binary installed above.
• Headless never starts the MV3 service worker in this setup — run headful (headless: false); windows open briefly at an offscreen position.
• The extension's storage can only be inspected/seeded through the service worker target (browser.waitForTarget(t => t.type() === 'service_worker'), then target.worker().evaluate(...)).
• Seed data, then relaunch with the same userDataDir so startup paths (e.g. storage migrations) run against the seeded state.

Run

node bench.mjs <path-to-extension-build> fixed     # perf + 16 functional checks
node bench.mjs <path-to-old-build>       baseline  # perf numbers only, for A/B

The fixed mode asserts: legacy v1 blob migrates to per-domain v2 keys + index; typing is captured and coalesced; shadow DOM (incl. nested) and contenteditable capture works; <select> change captures; password values are never stored; a settings push blocks new captures; the popup lists, filters and renders domains. Results land in result-<mode>.json + popup.png.

Perf regression bar

With a ~7.6 MB stash (400 domains), typing 25 bursts in one tab must produce 0 ms of measured main-thread blocking in a second open tab. The v1 architecture measured ~1100 ms blocked / 5 long tasks in that scenario (the storage.onChanged broadcast storm). If the probe tab shows blocking again, someone re-introduced a content-script storage.onChanged listener or a whole-stash write on the capture path.