Fix flaky browser tests in CI

[coji]

Problem

Browser tests occasionally fail in CI with:

Error: Browser connection was closed while running tests. Was the page closed unexpectedly?

Observed behavior

• concurrency.test.ts passes
• durably.test.ts fails with WebSocket connection closed
• Happens intermittently (flaky)

Possible causes

1. OPFS resource contention between parallel test files
2. SQLocal cleanup not completing before next test starts
3. Memory pressure in CI environment
4. Test file execution order issues

Investigation needed

• Review tests/helpers/browser-dialect.ts OPFS usage
• Check cleanup patterns in afterEach hooks
• Consider fileParallelism: false as workaround (but impacts speed)
• Look into SQLocal's known issues with concurrent access

References

• Observed in PR feat: add generic type support for getRun/getRuns and useRuns #15 CI runs

[coji]

Investigation Summary

After thorough investigation, the reported browser test flakiness cannot currently be reproduced:

Findings

1. CI Logs Analysis: Reviewed 15+ failed CI runs from the past weeks. All failures were due to:

   • Format/lint errors (prettier check failures)
   • No "Browser connection was closed" errors found in any run logs

2. Recent CI Status: All recent CI runs (past 20+) are passing, including browser tests

3. Local Testing: Ran browser tests 3 times consecutively - all passed (118 tests, 7.6-7.8s each)

4. PR feat: add generic type support for getRun/getRuns and useRuns #15 Status: The PR referenced in this issue passed all checks successfully

Current Mitigations Already in Place

• vitest.browser.config.ts already has retry: 2 configured
• Each test uses unique database names (via counter + timestamp) for isolation

Possible Explanations

1. Transient Infrastructure Issue: The original failures may have been caused by GitHub Actions runner instability
2. Already Fixed: Recent changes may have inadvertently resolved the issue
3. Timing-Dependent: The issue may require specific timing conditions that are hard to reproduce

Recommendation

Monitor CI runs for the next few weeks. If the "Browser connection was closed" error reoccurs:

1. Capture the full CI log for analysis
2. Consider fileParallelism: false as a workaround (trades ~2x test time for stability)
3. Investigate SQLocal's BroadcastChannel usage for potential race conditions

Keeping this issue open for monitoring. Will close if no recurrence after 2 weeks.

[coji]

追加調査 (2026-01-04)

ローカル再現テスト

ローカル環境で50回以上連続でブラウザテストを実行しましたが、「Browser connection was closed」エラーは再現できませんでした。

• 通常実行: 30回連続成功
• CPU負荷をかけながら実行: 20回連続成功

関連情報

Vitest の Discussion で同様の問題が報告されています：

• What is the best way to debug `Browser connection was closed while running tests`? vitest-dev/vitest#7981

原因: Chromium の新しいヘッドレスモードが不安定な場合がある
解決策: channel: 'chrome' を設定してシステムの Chrome を使用する

provider: playwright({
  launchOptions: {
    channel: 'chrome',
  },
})

現状の判断

• ローカルでは再現不可
• CIでも最近は発生していない（12/21以降、ブラウザテスト起因の失敗なし）
• retry: 2 が既に設定されているため、万が一発生しても自動リトライされる

次のステップ

再発した場合に channel: 'chrome' の設定を適用することを検討。現時点では様子見とします。