Investigate Desktop E2E failures observed in PR #5316 CI

[liugddx]

Status

Side Chat admission race reproduced with a controlled real-Host response gate. Upstream #5315 merged the same synchronization fix during investigation. PR #5316 now includes that main commit via merge 8a3c12ea190bd62d68a961ad900d8a4ceae69f86; Linux CI passed, including 39 Desktop E2E tests, at https://github.com/apache/maka/actions/runs/34941877780.

Draft #5340 was closed as redundant with #5315. The reported CI blockage is resolved on the updated PR. The Skill Draft menu-click failure did not recur in the hosted run; its independent root cause was not established. Reopen with a fresh trace if that specific failure recurs.

Original failures

Run https://github.com/apache/maka/actions/runs/34922439867:

• Side Chat, side-chat-followups.spec.ts:155: expected successor two in the queue; the screenshot shows the text still in the editor.
• Skill Draft, skill-draft-lifecycle.spec.ts:59 called from line 117: the root HTML element intercepted the click on the Project Only option, then the option detached. This was not a click on Cancel.

Confirmed Side Chat ordering

A temporary Electron probe withheld the real Host submit response after admission. The first queue row appeared while the Composer Send control remained disabled. A second Enter did not cause another Host invocation and preserved the second draft. Releasing the response, waiting for the enabled Send control and pressing Enter admitted the second message.

All probe behavioral assertions passed. Windows teardown subsequently failed deleting Chromium's DIPS file with EBUSY, so the overall probe run was not green. The temporary probe was not added to the permanent E2E inventory.

#5315 now uses the existing awaitSendReady helper with the companion scope before consecutive sends. It preserves the assertions and does not add sleeps or retries. No further duplicate fix is required for that race.

Validation and limits

• Clean Windows install completed using Node 24.19.0 / npm 11.19.0 and the official registry.
• Unmodified main 72cd8b1f5 and PR 27b39b2f6 baselines both passed the two Skill Draft cases. Side Chat passed the original hosted failure phase but later failed locally (overall timeout / reconnect evaluation).
• After merging main 1ae4d5b89 into feat(desktop): provide managed HTML artifact preview endpoints #5316: clean full workspace build passed; all 56 managed-preview/Artifact IPC/native-capability tests passed.
• Post-merge Windows E2E: the original failing Skill Draft cancellation case passed. Side Chat and the other Skill Draft case exhausted the existing 60-second whole-test budget. These are different failure modes from the original hosted report; no timeout was raised.
• Local traces/screenshots are retained. Linux CI is still required; a Windows pass does not prove that an intermittent Linux menu-click failure is fixed.

The earlier claim that unchanged test files proved these failures unrelated to #5316 was too strong and is withdrawn. Side Chat now has a demonstrated ordering defect in the test; Skill Draft still needs causal evidence if it recurs on Linux.

Related: #5316, #5315, #5340

[liugddx]

Updated investigation: latest upstream/main is 2bc8a54. Both failing E2E files remain unchanged relative to PR #5316 and main. I attempted a clean worktree run, but the worktree lacks the workspace-built @maka packages and Playwright dependencies, so the tests cannot run without the repository install/build pipeline. The hosted failure remains environmental/timing-related until reproduced with the CI setup; no product regression is established yet.