bmuxKeep local verification to typecheck, lint, and unit tests for the changed feature. Run affected GUI tests in GitHub Actions or Tart. The full suite runs on main in CI; it does not need to run on the working desktop for every edit.
Select the affected behavior
Search existing assertions when changing a control, label, selection rule, or shared test helper. Include its existing callers and relevant states, such as blank and loaded panes. A newly added test alone may miss an existing route.
For example, changes to pane profile controls affect profile selection, opening settings, and proxy settings. A useful pattern is:
gh workflow run check.yml -R tonisives/bmux --ref YOUR_BRANCH \
-f test_pattern='profile icon|pane address shows profile|saved proxy settings|proxy toggle|profile proxy settings|blank pane can change profile|new sessions use the saved preference'
For a reproduced timing failure, repeat just its affected cases four to eight
times with -F repeat_each=8. Repetitions expose failures; they do not retry a
failed assertion or turn a failing run green.
Verify the commit being merged
Commit and push the implementation branch before dispatching. Incorporate main before the final verification if main changed behavior covered by these tests. Wait for the selected run to finish:
gh run list -R tonisives/bmux --workflow Check --branch YOUR_BRANCH
gh run watch RUN_ID -R tonisives/bmux --exit-status
gh run view RUN_ID -R tonisives/bmux --json headSha,conclusion,url
git rev-parse HEAD
The tested SHA must match the implementation commit being merged. Record the run URL and selected tests in the task result. A queued run, a run on an earlier commit, or a green Portable hosts run does not verify the macOS GUI cases.
Diagnose a failure
Check the first failed action and download electron-test-results. Profile and
command tests save native focus and mouse delivery records on failure. For their
DOM action traces, dispatch the affected cases with -F trace=true; failed traces
are saved as trace.zip. Tracing is off by default because recording snapshots
adds work and can change the timing of a race. Confirm a timing fix with tracing
off.
Wait for observable readiness: the intended pane is selected, a newly mounted control has finished taking focus, navigation has committed, or a native view is attached. Avoid fixed sleeps, forced DOM clicks, and blanket timeout increases. Keep test-owned state and cleanup independent of preceding tests.
Findings from the September 30 profile regression
f2ddc6a moved profile controls from the status bar into pane address bars.
Opening settings in a blank pane then required a picker with an autofocus select.
While a blank pane initializes, its profile control can change from the settings
route to the profile picker route. The helper chose between those routes using an
immediate label count, before clicking. The failed snapshot still showed the picker
and URL prompt, with no settings dialog. A trace also showed the chosen button
label disappearing before the click. The same case failed in several later main runs.
An isolated CI repetition reproduced it five times out of eight, so earlier tests
were not required to trigger the failure.
The helper now matches both labels, clicks once, and waits for the dialog or picker that actually opened. It also waits for the picker's autofocus before clicking settings. Assertions about the selected profile, saved proxies, and credentials remain in place.
Earlier failures had different causes: shared recovery profile state, unbounded Electron shutdown, and native window focus or remote signaling deadlines. Fixing one case did not validate later changes to a different UI route. Keep verification bound to the changed behavior and the exact commit.