You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Flaky e2e: TestJoinHistoryFolderPlain finds no assistant entry in session_joined history #392
Relay main push CI at 6730e9f (run 38444948647, e2e) failed: g1_history_folder_test.go:113: session_joined history has no assistant entry with the reply; history: [{"role":"user","content":"hello history"}]
The same tree passed the whole e2e job in PR #390's last run. Run alone on the test machine, TestJoinHistoryFolderPlain and TestJoinHistoryFolderWithSpaceAndDot pass 20 of 20 at 6730e9f and at 7c3c45c (go test -race -count=10). It shows only under the full parallel e2e load.
Root cause
Not found in the 30-minute pass. The fake claude writes the user and assistant transcript lines before it reports the result (e2e/fakes/fakeagent/claude.go:79), and the test rejoins after the result, so the history should hold both. The user line was there and the assistant line was not. A partly flushed transcript read, or a history merge that drops the assistant entry under load, would both fit; neither is shown.
Fix
Next: reproduce under load (-count with -parallel 32 and the rest of ./features), then fix the wait or the read.
Closing as stale. These flakes predate the stack-per-run harness (devboxWorld#25, merged in barelyworkingcode/devboxWorld#32 on 2026-10-17): each verification run now gets its own seeded stack, so shared-world load and mid-run rebuilds no longer reach it. Owner decision, 2026-10-18. If it fails again on a stack run, reopen with the run log.
Activity
barelyworkingcode commentedon Oct 13, 2026
Rerun of the failed e2e job on main (run 38444948647) started for this flake.
barelyworkingcode commentedon Oct 19, 2026
Closing as stale. These flakes predate the stack-per-run harness (devboxWorld#25, merged in barelyworkingcode/devboxWorld#32 on 2026-10-17): each verification run now gets its own seeded stack, so shared-world load and mid-run rebuilds no longer reach it. Owner decision, 2026-10-18. If it fails again on a stack run, reopen with the run log.