Verify each turn end to end, with a recorded test run

Off by default. Turn it on (or back off) in Global settings → Auto QA (the default for new sessions), per session in its Session settings or in the New Session form (Settings default / On / Off), from the switch in the Auto QA pane, or with sessionboxer new --no-e2e / --e2e.

With it on, every turn the agent finishes is followed by a hidden verification turn. The agent runs the e2e-verification skill installed in the box: it looks at what the turn changed (git status in each repository, what it did), and either records the run as skipped with a reason (an answer, research, nothing testable) or plans 2–5 test cases from your prompt and what it understood you wanted (up to 10 for a very large change, rarely). It then starts a desktop recording and runs the cases one by one with the mouse, keyboard and browser of the box, captioning the video as it goes. A case that fails is fixed — that is normal agent work, in the same turn — and rerun as a new cycle of the same case, at most three fix attempts, after which it stays failed. At the end the recording stops and the agent replies with the video, which plays inline in the chat like any recording.

You can also ask for one whenever you like: Run now in the Auto QA pane starts the same verification turn for the work so far, measured against your last request, whether the switch is on or off (the agent still records it as skipped when there is nothing to exercise). The button waits while the agent is working, while a run is already open, or while the session is stopped.

The Auto QA pane (Verify on the phone) opens by itself the moment the first case starts running, for the session you are looking at. It shows the run's status, a progress bar, each case with its state, a live timer while it runs, its cycle, the agent's note and a screenshot, the steps and expected result on click, the video when done, and earlier runs under Earlier runs. A marker at the end of the turn in the chat ("Verified: 4/4 passed · 2:13 · video", or "Verification skipped: …") opens the pane on that run. A run stays with the session across Stop/Resume; stopping the session or the Control Plane while one is open marks it aborted.

A verification turn never verifies itself, and a queued message waits until the verification is over. Each verified turn costs a second turn of model time plus the minutes the agent spends driving the desktop; switch it off for sessions where that is not worth it.