e2e (linux) and e2e (windows) jobs flaky/broken on dev — ~98% fail today

[YumaKakuya]

Description

test workflow's e2e (linux) and e2e (windows) jobs have been flaky-to-broken on dev for several days, and ~98% fail today (2026-04-11). The failing runs always show the same tail in the job log: session.created → message.updated → message.part.updated → default disposing all instances, then a 30-minute gap, then The action 'Run app e2e tests' has timed out after 30 minutes.

Rough pass rate on dev over the last few days (counted from gh api /actions/workflows/.../runs?branch=dev):

• 2026-04-07: ~40% pass
• 2026-04-08: ~12% pass
• 2026-04-09: ~7% pass
• 2026-04-10: ~13% pass
• 2026-04-11 (through ~15:00 UTC): ~2% pass (≈ 3 successes in ~150 runs)

Many independent PRs hit the same tail — for example #22041, #22035, #22026, #22025, #22022, #22016 are all unrelated fixes but all fail the same way, and the dev push runs fail too. That makes it look like something in the test setup or session lifecycle rather than anything PR-specific.

Also, gh run rerun reports run cannot be rerun; its workflow file may be broken, so PRs currently have no way to retry a failing e2e job except by pushing a new commit.

Plugins

No response

OpenCode version

dev (HEAD 27190635ea)

Steps to reproduce

1. Open the test workflow's dev-branch runs: https://github.com/anomalyco/opencode/actions/workflows/test.yml?query=branch%3Adev
2. Pick any recent e2e (linux) or e2e (windows) job
3. Scroll to the tail of the job log and observe default disposing all instances followed by the 30-minute timeout

Screenshot and/or share link

No response

Operating System

GitHub Actions runners (ubuntu-latest and windows-latest)

Terminal

No response