QUICK ANSWER
Repeat the original reproduction after the code change and record the visible result. Ask the agent to return the watch link with the actions exercised, test results, and remaining gaps. A focused proof video should show the starting state, the changed action, and the expected final state rather than a general tour of the app.
What should the verification recording show?
Turn the original failure into a small verification plan. If the bug was a filter resetting after Save, start with the filter selected, make the save action, and show the persisted selection. Use the same route and relevant state as the original take.
Keep unrelated setup out of the recording. A video that starts after the successful action cannot show whether the action produced the result. Show enough of the sequence for a reviewer to check the causal claim.
How do you ask an agent for proof?
Give it a direct request after the fix:
Repeat the reproduction from this recording against the changed code. Record the visible result and return the Clipy link. State the steps exercised, the tests run, and anything you could not verify. Add the link to the PR description.
Use automated recording for a driven browser verification. For an existing native window or test video, follow the proof workflow and its source requirements.
What makes a mark trustworthy?
A timed note connects an observation to a moment in the take. Name which check ran and what it observed; Clipy keeps live-page assertions and driver notes with separate provenance.
Clipy supports live-page assertions for browser sessions, so a reviewer can tell a passed automated check from a narrated observation at a glance.
What should happen if verification fails?
Keep the failure visible in your task status and investigate it. The session run wrapper discards takes when the driver exits non-zero, so a failed scripted verification does not become an uploaded success recording.
If you need to retain the failed reproduction as evidence, use an explicit recording session and explain the result. Do not ignore a failed assertion or change its exit code to make the take look successful.
How should the PR describe the result?
Lead with the trigger and resulting behavior. For example: 'Saving an Active filter now preserves the selection. Verified by repeating the original reproduction and checking the persisted state after reload.' Link the recording beside that claim.
Use the recording-to-PR-description tool for a draft if helpful, then check the text. Keep actual test results separate from a manual recording, and state any paths that remain untested.
What does a proof video show?
Exactly what was observed in that run, ready to sit beside your tests and the code. Return to the original bug recording when the intent or reproduction becomes unclear.