[Rizky92]

This PR attempt a fix to mode selector. Per claude code v2.1.301, the mode selector would change mid-prompt. This was caused by a visual change to claude code's status bar, unable to recognize the current mode.

[oleg68]

Detector change looks good: gating the extra lines on the ⏸/⏵ glyph keeps menu text from being mistaken for the footer, and the negative test covers that. Two requested changes inline (stuck state on unchecked exception, flaky test). Also optional: sendShiftTab() (manual single Shift+Tab) still sets modeSwitchInProgress = false in its finally and can clear the flag while the worker is running; and please check the pom.xml version bump against current main after rebase.

[oleg68]

Thanks for the fix, the single-worker retargeting is the right design.

One robustness issue: the old finally { modeSwitchInProgress = false; } was removed, and this catch only handles IOException | InterruptedException. If sendOneShiftTab() throws an unchecked exception (e.g. an NPE because connector became null when the session was stopped mid-switch), editModeSwitchTarget stays non-null forever. After that every sendShiftTabsUntilMode call just "retargets" and returns (no worker ever runs again), and modeSwitchInProgress stays true, so the screen→model mode sync in the poll loop is disabled for the rest of the session.

Catching RuntimeException here reuses the existing cleanup block. (I'd avoid moving the cleanup into finally: on a normal exit it could clobber the target of a new worker started right after the loop ended.)

Suggested change

	} catch (IOException | InterruptedException ex) {
	} catch (IOException | InterruptedException | RuntimeException ex) {

[oleg68]

This test can be flaky: it assumes the second request arrives while worker 1 is still running, which only holds because the fake poll interval is 1 ms and the request follows immediately. On a loaded CI machine worker 1 may finish first, then a second thread is started and senders.size() is 2.

Make it deterministic by blocking the first Shift+Tab until the second request has been issued, e.g.:

java.util.concurrent.CountDownLatch secondIssued = new java.util.concurrent.CountDownLatch(1);
// in the overridden sendOneShiftTab():
senders.add(Thread.currentThread());
try { secondIssued.await(2, java.util.concurrent.TimeUnit.SECONDS); } catch (InterruptedException e) { Thread.currentThread().interrupt(); }
ccMode.updateAndGet(i -> (i + 1) % cycle.length);
// in the test body:
c.onEditModeComboChanged(EditMode.AUTO);
c.onEditModeComboChanged(EditMode.PLAN);
secondIssued.countDown();

[oleg68]

Minor / optional: claude-launch-tests/test_edit_mode_footer.py is a capture script without assertions. shutil.which('claude') returns None when Claude is not on PATH, which then fails with a confusing error inside subprocess.Popen; consider an explicit sys.exit('claude not found on PATH').

[oleg68]

Replaced with #195