Depot runners losing connection mid-job evict merge-queue entries #1857

Since ~20:15 UTC on 2026-10-29, Depot-hosted jobs in CI keep dying with
The self-hosted runner lost communication with the server (annotation on the
job; no log is retained). Still happening at 21:37 UTC.

Merge queue impact: #1819 was evicted twice, both times only by this:

• merge_group run 40947334054 (20:15): dispatch-tests, coverage (1), coverage (2) lost their runners exactly 10 min after start.
• merge_group run 40950997135 (20:39): 9 e2e legs plus hosted-e2e lost their runners at 20:54:50-51, all in the same second, on separate VMs.

PR impact (same annotation): runs 40944958480 (#1849: lint-ecosystems, dispatch-tests, coverage (2), e2e-build), 40949170855 and 40951029272 (#1698), 40951557857 (#1854), 40955713991 (#1849, coverage (2) at 21:37).

Why I think it's the platform, not the repo: many different job types
die (clippy, lint, coverage, e2e), several separate VMs drop in the same
second, and the jobs that die run the same code that passes on other
runners in the same window (e.g. runs 40955083994, 40957229892).
The last merge to main before this, #1828, only changes how bundler is
installed in the e2e jobs, and it can't affect clippy or coverage jobs.

Action: this is outside the repo. If it continues, check Depot status or
the runner pool. Re-adding a PR to the queue while this is ongoing will
probably get it evicted again.

Generated by Claude Code