test (macos-latest) is cancelled after ~15 minutes without ever acquiring a runner #890

test (macos-latest) is cancelled after ~15 minutes without ever acquiring a runner

Blocking #889, which is a README + one test file and touches no workflow.

What happens

Three attempts on the same commit, each started by gh run rerun --failed:

Every one ends conclusion: cancelled with zero steps and an empty log:
the job never started. While queued the API reports "runner": null; after
cancellation, "runner": "".

gh pr checks renders this as fail, which is what makes it expensive — it
reads as a test failure on a pull request that changes no code, and the first
instinct is to go looking for a broken test that never ran.

It is not the workflow, the label or the plan

A control, same job name, same workflow, same day — run 41845361857 on main,
14:13 UTC:

{"name": "test (macos-latest)", "conclusion": "success",
 "runner": "GitHub Actions 1000020451",
 "started_at": "14:14:30Z", "completed_at": "15:03:37Z"}

It acquired a runner and ran for 49 minutes — well past the ~15 minutes
that kills the queued ones. So macos-latest resolves, the account can obtain
macOS runners, and a job that gets one completes normally.

Ruled out in .github/workflows/ci.yml:

• no concurrency: block, so nothing in this repository cancels an in-flight
  run when another starts;
• no timeout-minutes: on any job (8 runs-on: entries, zero
  timeout-minutes:).

Why this is filed rather than retried again

Attempts 2 and 3 were cancelled at 15m01s each, to the second. Identical
durations are not the signature of a busy pool; they are the signature of a
fixed cap being reached. Retrying a reproducible cutoff a fourth time would be
treating a deterministic outcome as an intermittent one.

What would close it

Find what imposes the ~15 minute acquisition limit — an org or account runner
policy, a macOS concurrency ceiling, or a cancellation from outside this
repository — and either raise it or make the lane report "no runner" distinctly
from "the tests failed". The second half matters on its own: a check that
cannot say nothing ran costs a reader the same attention as a real failure,
every time.

Not attributed here: ~25 sessions share this account, and naming one as the
canceller without a control would be a guess.

🤖 Generated with Claude Code