[Mikola Lysenko (mikolalysenko)]

Fixes #1745

Problem

ci-ok runs one 2 s python3 check over toJSON(needs), alone, after every other CI job. It ran on depot-ubuntu-24.04-4, so each run paid a Depot host boot right on the merge's critical path. From #1745 (n=21, 10-27 13:07–20:16 UTC): start gap (job start → first step) mean 32 s, p90 131 s, max 155 s; since 16:17 the mean was 75 s.

Re-checked on the latest 4 runs before this change:

run	event	runner	start gap
40577778835	pull_request	depot-71zsq7z22d	32 s
40575163903	pull_request	depot-r3sr8hpl78	28 s
40574951881	push	depot-qbxw5fws87	5 s
40573814234	merge_group	depot-5mz2mw730c	123 s

Change

• .github/workflows/ci.yml, job ci-ok: runs-on: ${{ vars.CI_OK_RUNNER || 'ubuntu-latest' }} (GitHub-hosted by default) instead of the Depot/DISABLE_DEPOT_RUNNERS expression. Setting the repo variable CI_OK_RUNNER (e.g. depot-ubuntu-24.04-4) moves it back without a merge. Name, if: always(), needs, timeout-minutes and the check are unchanged.
• scripts/tests/test_ci_scheduling.py: new test_ci_ok_defaults_to_a_hosted_runner pins that runs-on line.

Hosted queue check (asked for in #1745, because #1372 once moved ci-ok off ubuntu-latest during a 9–18 min hosted backlog)

GitHub-hosted jobs in merge-queue-fail-fast, merge-queue-janitor and pin-check, up to 3 runs per workflow per day, 10-24 → 10-27 (n=27): queue wait (created → started) p50 2 s, p90 2 s, max 3 s; start gap (started → first step) p50/p90/max 1 s. Well under the 10 s p90 bar. If hosted Linux ever backs up again, set CI_OK_RUNNER; no PR or merge needed.

Expected saving

• Merge-queue critical path and PR feedback: ~0.5 min mean per run (32 s → ~3 s), 2–2.5 min per run during Depot slow-start episodes (CI perf (settings): slow Depot 16-core hosts since 10-13 15:20 add ~1.5–2.5 min to the merge-queue critical path (~0.6k Linux job-min/day) #1512).
• ~0.5 Depot job-min per run, ~25–50 Linux job-min/day moved to hosted (free for a public repo).

Measured result

This PR's CI run 40578824535 (head e631e18, draft, so the heavy jobs skip as designed): ci-ok job ran on hosted ubuntu-24.04 (GitHub Actions 1002782885). Queued 2 s, first step 0 s after start, check took <1 s, whole job 3 s, result success.

	baseline (#1745, Depot, n=21)	latest 4 Depot runs	this PR (hosted)
start gap	mean 32 s, p90 131 s, max 155 s	32 / 28 / 5 / 123 s	0 s
job duration	~5–159 s	6–128 s	3 s

The merge_group saving shows only after merge (the queue runs the base branch's workflow plus this change). The profiler will verify it against the 10-27 baseline.

Where each test runs

No test moves. ci-ok keeps its name and needs, so the required check is the same.

Risk

Low. Only the runner of the aggregate check changes. The workflow change takes effect on this PR's head commit; merge_group effects show after merge, and the profiler will verify them.

Validation: python3 -m unittest discover -s scripts/tests (374 tests OK); YAML parses; zizmor findings on ci.yml identical to main (59, 0 medium/high); actionlint 1.7.7 reports only pre-existing anchor/matrix findings, none in ci-ok.

Code review (/code-review high, before leaving draft)

Fixed in 800ff09: (1) a hard-coded hosted label meant a hosted backlog could only be undone by a PR that itself waits on ci-ok, so vars.CI_OK_RUNNER is the escape hatch; (3) ubuntu-latest instead of ubuntu-24.04, matching every other hosted Linux job; (4) the assertion now lives in its own test; (6) dated measurements dropped from the YAML comment.
Not fixed: (2) hosted concurrency shared with the Windows/macOS legs under merge-queue peaks is a real unknown, but CI_OK_RUNNER now covers it and the profiler checks merge_group after merge; (5) moving other short critical-path jobs such as release-readiness off Depot is out of scope for #1745 and would need its own measurement.

https://claude.ai/code/session_011f9aePipa5qoK8iqhBVEbW

---

Note

Low Risk
Only the runner for the final aggregate status job changes; required-check logic and job dependencies are unchanged, with an optional repo variable to revert runner choice without a code change.

Overview
Moves the merge-blocking ci-ok aggregate job off Depot by default so the last step of every CI run no longer pays a Depot runner cold-start on the critical path (#1745).

ci-ok now uses runs-on: ${{ vars.CI_OK_RUNNER || 'ubuntu-latest' }} instead of the DISABLE_DEPOT_RUNNERS / depot-ubuntu-24.04-4 expression. The job name, needs, if: always(), and the short Python verdict step are unchanged; only the runner label changes. Workflow comments document that CI_OK_RUNNER (e.g. depot-ubuntu-24.04-4) is the repo-variable escape hatch if hosted Linux queues up.

scripts/tests/test_ci_scheduling.py adds test_ci_ok_defaults_to_a_hosted_runner to lock in that runs-on line.

^{Reviewed by Cursor Bugbot for commit 800ff09. Configure here.}

---

Generated by Claude Code

[Mikola Lysenko (mikolalysenko)]

bugbot run

---

Generated by Claude Code

[cursor]

Stale Bugbot comment from a previous run.

[Mikola Lysenko (mikolalysenko)]

bugbot run

---

Generated by Claude Code

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 800ff09. Configure here.}

[Mikola Lysenko (mikolalysenko)]

Burn-down agent: labeled Ready for review at head 800ff09.

• CI: 83/83 checks on 800ff09 passed or skipped (33 success, 50 skipped by path filters). Mergeable with main; no main-wide failures.
• /code-review high: 6 findings. Fixed 4 in 800ff09 (repo-variable CI_OK_RUNNER escape hatch, ubuntu-latest for consistency, own test, no dated numbers in YAML). 2 left as design notes in the description (hosted concurrency at merge-queue peaks; other short Depot jobs out of scope).
• Bugbot: reviewed 800ff09, no new issues. No open review threads.
• Reviewer: the one-line runs-on change in ci-ok is the whole effect; the merge_group saving only shows after merge.
• Slack: not announced, because this run had no Slack send tool.

---

Generated by Claude Code