Open-source AI browser agent for Chrome and Firefox (monorepo) π§
Manual test: screenshot redaction (issue #312)
Redaction needs a live vision model call and a real screenshot pipeline β
test/run.js only covers the pure helpers (selectRedactionRegions,
mapRegionsToImage, mergeRedactionFrameRegions). Use the fixture page below
to eyeball the actual pixelation before merging changes to
agent/screenshot-redaction.js, content/redaction-regions.js, or the
st.redaction.* settings UI.
Setup
- Serve the fixtures folder instead of opening it via
file://(avoids the "Allow access to file URLs" extension permission entirely):
then opencd test/fixtures && python3 -m http.server 8912http://localhost:8912/screenshot-redaction-manual.html. - Load the unpacked extension (Chrome:
chrome://extensionsβ Developer mode β Load unpacked βsrc/chrome. Firefox:about:debuggingβ This Firefox β Load Temporary Add-on βsrc/firefox/manifest.json). - Settings β Advanced β turn Screenshot redaction on.
Test 1 β Redaction on, take a screenshot
With the fixture page open and active, ask the agent to /screenshot (or
just "take a screenshot" / "verify the page"). In the returned image, check
every element against the REDACT/SKIP tag printed next to it on the
page:
- Password, email, phone, plain text, search, url, number inputs, the
textarea, the contenteditable div, and the
<select>β pixelated boxes. - Checkbox, radio, range, color, submit button β untouched (never collected).
- The two standalone "Page text β detected PII patterns" lines β pixelated.
- The "Page text β heuristic should skip" lines (the bare year, the email embedded in a full sentence) β untouched, by design.
- The nested same-origin iframe's password/email/phone fields β pixelated
and correctly aligned to their on-screen position (confirms
mergeRedactionFrameRegionsframe-to-frame coordinate mapping).
Test 2 β Known, documented gaps (should still leak)
These are expected to NOT be redacted β they're limitations called out in the Settings warning copy and the website FAQ, not regressions:
- The canvas-drawn PII text β DOM heuristics cannot see canvas pixels.
- The password field inside the
srcdociframe β Chrome'smatch_about_blankcontent-script matcher doesn't coversrcdocframes.
If either of these two ever becomes covered by a future change, update the
FAQ/settings copy (web/build/locales/en.json faq.screenshot_redaction,
and st.redaction.warning in the locale files) to stop describing it as a
limitation.
Test 3 β Toggle off (control)
Turn Screenshot redaction off, take another screenshot of the same page, and confirm nothing is pixelated β the password field, emails, and phone numbers are all fully legible. This is the regression check that the feature is truly opt-in and doesn't run at all when disabled.
Test 4 β Fail-open on a fresh navigation
Navigate to the fixture page and immediately (within ~1 second, before
document_idle content scripts have had time to inject) ask for a
screenshot. The known fail-open behavior means the screenshot may still be
sent unredacted rather than blocked β this is expected per the Settings
warning, not a bug to file.