Finish report and preview

A run that completes, meaning it ran out of work on its own, hands over two things: a finish report and a live preview of your app. Runs that were paused, stopped, halted or ran over budget do not.

When there is an app to show, a notice says so: The app is finished. Its report and preview are ready. Both are in the Manager’s Report tab, side by side.

The finish report

The report is built from the run’s own record, with no model call. It never re-runs your tests; it reads what the loop did.

• What was built: the files the run created or changed, the cards it finished, and each bubble’s last summary in its own words.
• How to run it: the command from your README that starts the app (npm run dev, npm start and so on), with the install step first when dependencies are missing.
• What was tested: the last test command the loop ran, its exit code and when.
• What is missing: anything that stops the run counting as finished, such as no README, no test script, tests never run, tests failing, cards still open, or a preview that crashed.

The report calls the app finished when:

1. the README gives one command to run it,
2. the last test run passed,
3. no bubble ended on FAIL or ERROR, and
4. the preview did not crash.

With a preview running, the report also takes screenshots of the app. Reports are saved in the project, in the app’s folder, as JSON for the tab and Markdown for your editor. The last 20 are kept.

The live preview

The preview is your app’s dev server, one per board. It is held apart from runs: a run ending does not stop it, and a new run does not stop it either, so a Vite app keeps following the loop’s edits.

• Start, restart and stop it from the Report tab, and read its log there.
• At the end of a run it starts on its own only when the board is in Full Auto and the command is allowed. Otherwise press Start.
• A board that sandboxes commands cannot run a preview.
• If the server dies, the tab shows its exit code and last lines instead of silently going blank.

The visual review

In a loop that builds a user interface, the reviewer runs a visual review. It opens the app in a real browser at 1280 × 800 and at 390 × 844 and checks each page.

• These fail it outright: an HTTP error, a dev-server error overlay, a blank page, a page wider than its window, a broken image, placeholder text such as lorem ipsum or TODO:, and stand-in image hosts.
• If you have set a vision model (Models page, Helper models), it also looks at the screenshots and judges them against the loop’s DESIGN.md. Layout problems, placeholders and broken images it marks as blocking fail the review; design and polish notes are advice.
• Without a vision model, the page checks alone decide, and the result says plainly that nobody looked at the screenshots.

The agents’ browser is an installed Chrome or Edge. If you have neither, Settings, Agent browser offers to download a Chromium, telling you its size first.