All pages

Close the loop · Lesson 1.2 · 12 min

Verify in the browser

Have the agent open the page, read the console and check what a user would see.

After this lesson you have the agent check its own front-end work the way a user would see it: it opens the page, reads the console and network log, asserts on what's on screen, and keeps a screenshot baseline that catches the change nobody meant to make.

Time: 12 minutes · Works on: macOS app. The cloud workspace has its own browser tools on a different engine.

Compiling is not working

For back-end code, a passing test suite is strong evidence. For anything a person looks at, it's weak. A component can type-check, render without throwing, pass its unit tests, and still show a blank card because an API call returned 401. The agent won't know unless it looks.

So make looking part of the job. A front-end task isn't done until the agent has loaded the page in a browser and checked three things:

  1. The console is clean. No new errors, no unhandled promise rejections. Warnings that were there before the change are fine. New ones need a reason.
  2. The network is clean. No 4xx or 5xx responses, no failed requests, for the flow you changed.
  3. The page shows what it should. The text, the element, the state you asked for, checked by an assertion rather than described from memory.

Then it shows you a screenshot. The screenshot is for you. The assertions are for the agent.

Assertions over eyeballing

A model looking at a screenshot will usually tell you it looks fine. That's the same problem as the agent grading its own work, with pixels. A check that passes or fails on its own is better:

  • "The heading reads Pricing."
  • "There are exactly three plan cards."
  • "After clicking Subscribe, the URL ends in /checkout."
  • "The Save button is disabled until the email field is filled."

Each of these is one call and returns a plain pass or fail with the actual value when it fails. Put them in the brief as done-when lines, the same way the previous lesson put a failing test there.

Baselines catch what you didn't ask about

Assertions check what you thought of. A visual baseline checks everything else. Save a screenshot of a page in a known-good state, then compare against it after each change. A shifted margin, a missing icon or a font that stopped loading shows up as a pixel difference, even though no assertion mentioned it.

Use baselines on a few pages that matter and change rarely. On a page under active redesign, every compare fails and the baseline becomes noise.

A worked example

You asked the agent to add a "Copy link" button to each row of a share dialog. Your brief ends:

Done when, at http://localhost:3000/share:
- each row has a button with the text "Copy link"
- clicking the first one shows the text "Copied"
- the browser console has no errors and no request failed
Show me a screenshot of the dialog when you're finished.

The agent makes the change, then opens the page in the built-in browser, counts the buttons, clicks the first, and waits for "Copied". That check fails: the clipboard call threw. It reads the console, finds NotAllowedError from the clipboard API, fixes the handler, and runs the checks again. Now they pass, the console is clean, and the screenshot shows the dialog. The bug was caught and fixed without you opening the page.

Do it in TRMNLY

  1. Put the app next to its terminal. Start your dev server in a pane, for example npm run dev. Show the right sidebar with ⇧⌘\, click Web, and open the server's URL in the address bar. The agent drives this same tab, so you can watch what it does.
  2. Point at what you mean. Click the crosshair beside the address bar, Design mode — click an element to hand it to the agent, and click the element. The agent gets its selector, styles, HTML and a cropped screenshot, so "the button" in your brief is unambiguous. The page's own click handlers don't fire while you pick.
  3. Have the agent read the console. browser_console returns the tab's captured console.* messages, uncaught errors and unhandled rejections. Ask it to filter to errors, and to drain the buffer before the check (clear: true) so it only sees what the change produced. You can read the same log yourself in the DevTools panel, Toggle DevTools (Console / Network / Elements).
  4. Have it read the network. browser_network returns each fetch and XHR call with method, URL, status and duration. With only_errors: true it keeps only failed, 4xx and 5xx requests, which is the check you want.
  5. Assert, don't describe. Ask the agent to use browser_expect. It retries until the condition holds or a timeout passes, so a check doesn't fail just because the page was still loading. It checks whether an element is visible or hidden, its text, an attribute, the URL, a count of matching elements, and whether a control is enabled or checked. For one-off text checks, browser_assert_text and browser_assert_url fail loudly with the actual value.
  6. Take the screenshot. browser_screenshot captures the viewport, one element by selector, or the full page. The image goes to a vision-capable model with the result, and the file is saved under ~/Library/Application Support/TRMNLY/agent/screenshots/ with its path returned, so you can open it. Keep the browser sidebar open while it runs: a collapsed sidebar can't be captured in a WebKit tab.
  7. Save a baseline. Ask the agent to call browser_visual_assert with a name and mode: "create" on a page you've checked by hand. That saves .trmnly/snapshots/<name>.png in the project. Later runs use the default assert mode and fail when the page differs beyond the tolerance. Pass a selector to compare one element instead of the whole viewport. Commit the snapshots folder if you want the baseline shared.
  8. Check other conditions. browser_viewport emulates a screen size so media queries recompute, and browser_emulate sets dark mode, reduced motion, timezone, locale or offline before the page loads. In a WebKit tab the viewport is approximated with zoom and a script, and screenshots show the sidebar's real size. Choose Chrome — TRMNLY profile from the engine chip when the layout viewport must really be the size you asked for.
  9. Make the flow repeatable. Ask the agent to record the check as a saved browser action with browser_action_record_start and browser_action_record_stop. It appears in More ▾ → Flows under Browser actions and replays with no model involved. To keep it as a permanent test, ask for browser_codegen, which turns the recording into a Playwright spec you can commit.

Check yourself

  • The agent reported the console and network results for your page without you asking.
  • At least one done-when line was checked with browser_expect or browser_assert_text, and you saw it fail once before it passed.
  • A baseline exists under .trmnly/snapshots/, and a deliberate CSS change makes its compare fail.
  • You can explain why a screenshot alone isn't proof the page works.

Go further